Remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement

Abstract

The invention provides a remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement. The remote sensing semantic segmentation algorithm comprises the following steps of S1, collecting two public remote sensing image semantic segmentation data sets; s2, training a proposed remote sensing image semantic segmentation algorithm according to the collected two public remote sensing image semantic segmentation data sets; and S3, detecting a to-be-detected remote sensing image according to the remote sensing image semantic segmentation model. The invention provides an innovative hybrid segmentation model, the model ingeniously combines an optimized convolutional neural network ResNet-50 and an advanced Swindow-Transform architecture, and innovatively integrates an RFB (Restrictive Field Block) feature fusion module and a WCCW (Window Connect Cross Window) feature extraction module. By means of the design, the problems that multi-scale feature correlation is insufficient and decoding layer features are not fully utilized are effectively solved, and therefore the accuracy of remote sensing image semantic segmentation is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement. Background Art

[0002] Traditional methods based on convolutional neural networks (CNNs) have made some progress in the field of image semantic segmentation. CNNs simulate the receptive field mechanism of the human visual system, extract local features of images using convolutional layers, and expand the receptive field through pooling layers to enhance the recognition of image structures. However, the locality of CNNs limits its ability to obtain global context information, which is particularly evident when dealing with small-scale, highly similar, and mutually occluded objects. These scenarios require models to capture more global context information and fine spatial features to improve the accuracy of semantic reasoning. In order to overcome these challenges, researchers have begun to explore new technical paths. Among them, the Transformer architecture has been introduced into the field of remote sensing image analysis due to its outstanding performance in natural language processing. Transformer effectively captures long-distance dependencies in images through the self-attention mechanism, providing new possibilities for solving small target detection and inter-class similarity problems in remote sensing images. In addition, the development of multi-source remote sensing image fusion technology improves the spatial and spectral resolution of images by integrating image data from different sensors and different resolutions, providing richer information for semantic segmentation.

[0003] In the research process of remote sensing image semantic segmentation, deep learning has become a powerful engine to promote the improvement of algorithm performance. Remote sensing image semantic segmentation based on deep learning can be divided into three categories: one is the method based on convolutional neural network (CNN). CNN automatically learns the hierarchical feature representation of the image through multi-layer convolution and pooling operations, effectively capturing the object boundary and texture information in the image. For example, the UNet model is widely used in remote sensing image semantic segmentation due to its success in medical image segmentation. Its symmetrical network structure includes two halves, feature extraction and upsampling, which can improve the boundary recognition ability, but CNN lacks the representation of long-distance pixel correlation and has a weak ability to extract global context information. Although AtrousSpatialPyramidPooling (ASPP) improves this, the computational complexity is high. One is the method based on Transformer. Transformer captures global dependencies through the self-attention mechanism, providing a new solution for remote sensing image semantic segmentation. It can perform pixel-level classification more accurately, especially when processing high-resolution images, but the high computational complexity of Transformer is not conducive to processing high-resolution remote sensing images. In addition, due to its high number of parameters, a large amount of training data is required, which may be difficult to obtain in remote sensing images. The last category of methods combines convolutional neural networks (CNNs) with Transformers, which can give full play to the advantages of both: CNN's efficient ability to capture local features and Transformers' excellent performance in modeling global contextual information. This fusion strategy has greatly improved the accuracy of semantic segmentation of remote sensing images. However, the network structure of this hybrid model may become more complex, which may not only increase the training and inference time of the model, but also cause problems with multi-scale feature correlation and insufficient utilization of decoding layer features. These problems need to be solved by carefully designing the network structure and optimizing the algorithm to ensure that the efficiency and scalability of the model are maintained while improving performance. Summary of the invention

[0004] The present invention provides a remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement, which solves the problem of multi-scale feature correlation and the problem of insufficient utilization of decoding layer features.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: The remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement includes the following steps: Step S1: Collect two public remote sensing image semantic segmentation datasets; Step S2: training the proposed remote sensing image semantic segmentation algorithm based on the two public remote sensing image semantic segmentation datasets collected above; Step S3: Detecting the remote sensing image to be detected according to the remote sensing image semantic segmentation model. In the above step S2, the remote sensing image semantic segmentation model is specifically: Parse the image data through the first ResBlock0 residual layer to obtain the feature map feature_map_0; Input the feature map feature_map_0 into the second ResBlock1 residual layer for parsing to obtain the feature map feature_map_1; Input the feature map feature_map_1 into the third ResBlock2 residual layer for parsing to obtain the feature map feature_map_2; Parse the feature map feature_map_2 through the fourth ResBlock3 residual layer to obtain the feature map feature_map_3; Parse the feature map feature_map_3 through the fifth ResBlock3 residual layer to obtain the feature map feature_map_4; The feature maps feature_map_1 and feature_map_2 are fused and parsed through the first Weight ChannelAttention layer to obtain the feature map feature_map_1_2CAF; The feature maps feature_map_2 and feature_map_3 are fused and parsed through the second Weight ChannelAttention layer to obtain the feature map feature_map_2_3CAF; The feature map feature is divided into small blocks through the Patch Embedding layer, and then fused with feature_map_1 and parsed through the first Swin Transformer Block 1 to obtain the feature map feature_map_S1; The feature map feature_map_S1 is downsampled through the Patch Merging layer, and then fused with feature_map_1_2CAF, and then parsed through the second Swin Transformer Block 2 to obtain the feature map feature_map_S2; The feature map feature_map_S2 is downsampled through the Patch Merging layer and fused with feature_map_1_2CAF and feature_map_2_3CAF, and then parsed through the third Swin Transformer Block 3 to obtain the feature map feature_map_S3; The feature map feature_map_S3 is downsampled through the Patch Merging layer, and then fused with feature_map_2_3CAF and feature_map_4, and then parsed through the fourth Swin Transformer Block 4 to obtain the feature map feature_map_S4; The feature map feature_map_S4 is fused with feature_map_S3 and the output feature map F1 is subjected to feature extraction and feature fusion by the second RFB module to obtain the feature map feature_map_S3_S4_F; The feature map feature_map_S3_S4_F is fused with feature_map_S2 and the output feature map F1 is subjected to feature extraction and feature fusion by the first RFB module to obtain the feature map feature_map_S2_F; The feature map feature_map_S4 is further subjected to feature extraction and feature fusion by the fourth WCCW module to obtain the feature map feature_map_W4; The feature map feature_map_S3 is connected with feature_map_S3_S4_F and feature_map_W4 in the channel dimension to obtain the feature map G, which is further extracted and fused through the third WCCW module to obtain the feature map feature_map_W3; The feature map feature_map_S2 is connected with feature_map_S2_F and feature_map_W3 in the channel dimension to obtain the feature map G, which is further extracted and fused through the second WCCW module to obtain the feature map feature_map_W2; The feature map feature_map_S1 is connected with feature_map_S2_F and feature_map_W2 in the channel dimension to obtain the feature map G, which is further extracted and fused through the first WCCW module to obtain the coarse feature map feature_map_W1; Input the feature maps feature_map_W1, feature_map_W2, feature_map_W3 and feature_map_W4 into the SSRM (Laplacian Pyramid) module, which will fuse feature information at different levels and output a fine feature map feature_map_all that integrates multi-scale information; The coarse feature map feature_map_W1 and the fine feature map feature_map_all are connected in the channel dimension and input into the detection head to obtain the final result.

[0006] The above RFB module is specifically: The feature map F1 is input into the convolution layer of the 1*1 convolution kernel to change the number of channels to reduce the number of parameters and calculations of the model and retain important feature information. Then the output feature map is divided into three branches: F2, F3, and F4. The feature map F2 is input into the convolution layer of the 1*1 convolution kernel to enhance the expressiveness of the features and output the feature map F2_1. The feature map F3 is input into the convolution layer of the 3*3 convolution kernel to capture more complex local features and output the feature map F3_1. The feature map F4 is input into the convolution layer of the 5*5 convolution kernel to capture a larger range of local features and obtain a wider range of contextual information and output the feature map F4_1. Figure F2_1 is input into the dilated convolution layer with a dilation rate of 1 to extract contextual information and output the feature map F2_2; the feature map F3_1 is input into the dilated convolution layer with a dilation rate of 3 to expand the receptive field and extract a wide range of contextual information, and output the feature map F3_2; the feature map F4_1 is input into the dilated convolution layer with a dilation rate of 5 to further expand the receptive field and extract a wider range of contextual information, and output the feature map F4_2; finally, the feature maps F2_2, F3_2, and F4_2 of the three branches are connected in the channel dimension; in this way, the features of different scales and different receptive fields can be fused to obtain a feature representation that integrates multi-scale information.

[0007] The above WCCW modules are specifically: Input the feature map G into the window partition module to divide the input feature map into multiple local windows and output the feature map G1; pass the feature map G1 through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and output the feature map G1_1. Finally, input the feature map G1_1 into the self-attention module to enhance the feature representation of the feature map and capture the global dependencies between different regions in the feature map, and output the feature map G1_2; Input the feature map G into the cross window partition module to divide the feature map into local windows and perform cross-window information interaction so that different windows can share information, and output the feature map G2; pass the feature map G2 through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and output the feature map G2_1; finally, input the feature map G2_1 into the self-attention module to enhance cross-window information interaction and extract hierarchical features, and output the feature map G2_2; Connect feature map G1_2 and feature map G2_2 in the channel dimension to enhance their feature representation, integrate global and local feature representations, and allow the model to simultaneously utilize fine-grained local features and large-scale global context information to improve the detection capability of small objects, and output feature map G_12; The feature map G_12 passes through the convolution layer with a 1*1 convolution kernel, then passes through the Batch Normalization layer, and finally inputs it into the DW convolution layer to further optimize the feature representation, improve the performance and computational efficiency of the model, and output the feature map G_all.

[0008] The present invention provides a remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement, which has the following technical effects: 1) This paper innovatively proposes an effective method to address the challenges of poor inter-class similarity and small target detection accuracy in remote sensing images. The method revolves around an innovative decoder structure, namely the multi-directional dynamic complement decoder, which adopts a dual strategy to integrate the top-down CNN-based (RFB) module with the top-down Transformer-based cross-window context exchange (WCCW) module. This design aims to enhance the network's ability to extract multi-directional features and capture hierarchical patterns, while also achieving a complementary fusion of local and global correlations, which significantly enhances the model's feature recognition ability; 2) This paper innovatively combines ResNet-50 with Swin Transformer to construct a new network architecture; RFB is customized to enhance the model's feature extraction and generalization capabilities in local areas, which helps to have a more detailed understanding of complex spatial patterns. On the other hand, the WCCW module uses a multi-head self-attention mechanism to enhance the model's cross-window information interaction and long-range dependency modeling capabilities. This comprehensive feature processing framework is crucial to solving the complex dynamic problems of remote sensing image segmentation. In addition, the SSRM Laplacian pyramid is introduced for feature fusion, which is a new component that can further improve the model's accuracy in segmenting small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the present invention; Figure 2 It is a schematic diagram of the overall structure of the remote sensing image semantic segmentation algorithm proposed in the present invention; Figure 3 yes Figure 2 Schematic diagram of the RFB module structure; Figure 4 yes Figure 2 Schematic diagram of the WCCW module structure. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following content will systematically and completely describe the specific technical solutions of the present invention in combination with the drawings provided according to the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0011] Embodiment 1: like Figure 1 As shown, an algorithm for semantic segmentation of remote sensing images includes the following steps: S1. Collect two public remote sensing image semantic segmentation datasets: ISPRS Vaihingen dataset and ISPRS Potsdam dataset. S2. Training the proposed algorithm based on the collected remote sensing image semantic segmentation dataset to obtain a remote sensing image semantic segmentation model; S3. Detect the image data to be detected according to the remote sensing image semantic segmentation model.

[0012] The structure of the remote sensing image semantic segmentation model in S2 is: Parse the image data through the first ResBlock0 residual layer to obtain the feature map feature_map_0; Input the feature map feature_map_0 into the second ResBlock1 residual layer for parsing to obtain the feature map feature_map_1; Input the feature map feature_map_1 into the third ResBlock2 residual layer for parsing to obtain the feature map feature_map_2; Parse the feature map feature_map_2 through the fourth ResBlock3 residual layer to obtain the feature map feature_map_3; Parse the feature map feature_map_3 through the fifth ResBlock3 residual layer to obtain the feature map feature_map_4; The feature maps feature_map_1 and feature_map_2 are fused and parsed through the first Weight ChannelAttention layer to obtain the feature map feature_map_1_2CAF; The feature maps feature_map_2 and feature_map_3 are fused and parsed through the second Weight ChannelAttention layer to obtain the feature map feature_map_2_3CAF; The feature map feature is divided into small blocks through the Patch Embedding layer, and then fused with feature_map_1 and parsed through the first Swin Transformer Block 1 to obtain the feature map feature_map_S1; The feature map feature_map_S1 is downsampled through the Patch Merging layer, and then fused with feature_map_1_2CAF, and then parsed through the second Swin Transformer Block 2 to obtain the feature map feature_map_S2; The feature map feature_map_S2 is downsampled through the Patch Merging layer and fused with feature_map_1_2CAF and feature_map_2_3CAF, and then parsed through the third Swin Transformer Block 3 to obtain the feature map feature_map_S3; The feature map feature_map_S3 is downsampled through the Patch Merging layer, and then fused with feature_map_2_3CAF and feature_map_4, and then parsed through the fourth Swin Transformer Block 4 to obtain the feature map feature_map_S4; The feature map feature_map_S4 is fused with feature_map_S3 and the output feature map F1 is subjected to feature extraction and feature fusion by the second RFB module to obtain the feature map feature_map_S3_S4_F; The feature map feature_map_S3_S4_F is fused with feature_map_S2 and the output feature map F1 is subjected to feature extraction and feature fusion by the first RFB module to obtain the feature map feature_map_S2_F; The feature map feature_map_S4 is further subjected to feature extraction and feature fusion by the fourth WCCW module to obtain the feature map feature_map_W4; The feature map feature_map_S3 is connected with feature_map_S3_S4_F and feature_map_W4 in the channel dimension to obtain the feature map G, which is further extracted and fused through the third WCCW module to obtain the feature map feature_map_W3; The feature map feature_map_S2 is connected with feature_map_S2_F and feature_map_W3 in the channel dimension to obtain the feature map G, which is further extracted and fused through the second WCCW module to obtain the feature map feature_map_W2; The feature map feature_map_S1 is connected with feature_map_S2_F and feature_map_W2 in the channel dimension to obtain the feature map G, which is further extracted and fused through the first WCCW module to obtain the coarse feature map feature_map_W1; Input the feature maps feature_map_W1, feature_map_W2, feature_map_W3 and feature_map_W4 into the SSRM (Laplacian Pyramid) module, which will fuse feature information at different levels and output a fine feature map feature_map_all that integrates multi-scale information; The coarse feature map feature_map_W1 and the fine feature map feature_map_all are connected in the channel dimension and input into the detection head to obtain the final result.

[0013] The RFB modules are as follows: The feature map F1 is input into the convolution layer of the 1*1 convolution kernel to change the number of channels to reduce the number of parameters and calculations of the model and retain important feature information. Then the output feature map is divided into three branches, namely F2, F3, and F4. The feature map F2 is input into the convolution layer of the 1*1 convolution kernel to enhance the expressiveness of the feature and output the feature map F2_1. The feature map F3 is input into the convolution layer of the 3*3 convolution kernel to capture more complex local features and output the feature map F3_1. The feature map F4 is input into the convolution layer of the 5*5 convolution kernel to capture a larger range of local features and obtain more extensive context information and output the feature map F4_1. The feature map F2_1 is input into the dilated convolution layer with a dilation rate of 1 to extract context information and output the feature map F2_2. The feature map F3_1 is input into the dilated convolution layer with a dilation rate of 3 to expand the receptive field and extract extensive context information and output the feature map F3_2. The feature map F4_1 is input into the dilated convolution layer with a dilation rate of 5 to further expand the receptive field and extract a wider range of contextual information, and the feature map F4_2 is output. Finally, the feature maps F2_2, F3_2, and F4_2 of the three branches are connected in the channel dimension. In this way, features of different scales and different receptive fields can be fused to obtain a feature representation F1_FUSION that integrates multi-scale information.

[0014] The WCCW modules are as follows: The feature map G is input to the window partition module to divide the input feature map into multiple local windows and output the feature map G1. The feature map G1 is passed through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and the feature map G1_1 is output. Finally, the feature map G1_1 is input to the self-attention module to enhance the feature representation of the feature map and capture the global dependencies between different regions in the feature map, and the feature map G1_2 is output.

[0015] The feature map G is input into the cross window partition module to divide the feature map into local windows and perform cross-window information interaction so that different windows can share information, and output the feature map G2. The feature map G2 is passed through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and the feature map G2_1 is output. Finally, the feature map G2_1 is input into the self-attention module to enhance cross-window information interaction and extract hierarchical features, and the feature map G2_2 is output.

[0016] The feature map G1_2 and the feature map G2_2 are connected in the channel dimension to enhance their feature representation, integrate the global and local feature representations, and allow the model to simultaneously utilize fine-grained local features and large-scale global context information to improve the detection ability of small targets, and output the feature map G_12.

[0017] The feature map G_12 passes through the convolution layer with a 1*1 convolution kernel, then passes through the Batch Normalization layer, and finally inputs it into the DW convolution layer to further optimize the feature representation, improve the performance and computational efficiency of the model, and output the feature map G_FUSION.

[0018] This invention not only solves the problem of missed detection in complex backgrounds and false detection caused by similarity between classes of traditional detection algorithms, but also significantly improves the generalization ability of the model by targeting the problem of sample imbalance. Through this method, we can more accurately capture the details of buildings in remote sensing images, providing strong technical support for urban planning, building monitoring and other fields.

[0019] Embodiment 2: In order to verify the effectiveness of this solution, further explanation is given below in combination with experimental data.

[0020] We use PyTorch to implement all experiments on a single NVIDIA RTX3.90 GPU and use Adam as our optimizer with a learning rate of 0.0006. We use SoftCrossEn-tropyLOSS and DiceLOSS as the joint loss function and use the backpropagation index to measure the gap between the 2D segmentation map and the ground truth we obtain. For each method, we use F1, mIOU and OA as evaluation indicators; To verify the effectiveness of our proposed model, our algorithm is evaluated against the SOTA method on a challenging benchmark dataset. Finally, we compare the model complexity with other models. We use F1, mIOU and OA as reference evaluation indicators to evaluate the algorithm effect. The parameter settings and experimental environment are consistent. The results are shown in Table 1: Table 1 Comparison with the most advanced algorithms on the Vaihingen dataset

[0021] From the experimental results in Table 1, we can see that the average F1, OA, and mIOU of our algorithm on the Vaihingen dataset increased by 1.75%, 2.12%, and 2.14%, respectively. This demonstrates the remarkable results of the interactive fusion method between CNN and Transformer, and emphasizes the advantages of the model in short-range and long-range complementary modeling. Our proposed model achieved good results for the two categories of targets, impervious pavement and cars, which proves that our model has made great progress in the segmentation ability of small-sized objects.

[0022] In order to verify the generalization and effectiveness of the algorithm proposed in this paper, we chose to conduct experiments on the POTSDAM dataset and compared it with the most advanced algorithms: Table 2 Comparison with the most advanced algorithms on the POTSDAM dataset

[0023] The improved algorithm of the present invention achieves excellent performance in the key performance indicators of remote sensing image semantic segmentation. Compared with other mainstream object detection networks, it shows significant advantages in many aspects. Specifically, the algorithm of the present invention achieves the best segmentation results in the six categories of impervious surfaces, buildings, low vegetation, trees, and cars. Its average F1, OA, and mIOU increase by 0.87%, 1.06%, and 0.77% respectively compared with the optimal algorithm. This undoubtedly proves the superior performance of our proposed model.

Claims

1. A remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement, characterized in that: The following steps are involved: Step S1: Collect two public remote sensing image semantic segmentation datasets; Step S2: training the proposed remote sensing image semantic segmentation algorithm based on the two public remote sensing image semantic segmentation datasets collected above; Step S3: Detecting the remote sensing image to be detected according to the remote sensing image semantic segmentation model.

2. The remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement according to claim 1, characterized in that: In step S2, the remote sensing image semantic segmentation model is specifically: Parse the image data through the first ResBlock0 residual layer to obtain the feature map feature_map_0; Input the feature map feature_map_0 into the second ResBlock1 residual layer for parsing to obtain the feature map feature_map_1; Input the feature map feature_map_1 into the third ResBlock2 residual layer for parsing to obtain the feature map feature_map_2; Parse the feature map feature_map_2 through the fourth ResBlock3 residual layer to obtain the feature map feature_map_3; The feature map feature_map_3 is parsed through the fifth ResBlock3 residual layer to obtain the feature map feature_map_4; The feature maps feature_map_1 and feature_map_2 are fused and parsed through the first Weight ChannelAttention layer to obtain the feature map feature_map_1_2CAF; The feature maps feature_map_2 and feature_map_3 are fused and parsed through the second Weight ChannelAttention layer to obtain the feature map feature_map_2_3CAF; The feature map feature is divided into small blocks through the Patch Embedding layer, and then fused with feature_map_1 and parsed through the first Swin Transformer Block 1 to obtain the feature map feature_map_S1; The feature map feature_map_S1 is downsampled through the Patch Merging layer, and then fused with feature_map_1_2CAF, and then parsed through the second Swin Transformer Block 2 to obtain the feature map feature_map_S2; The feature map feature_map_S2 is downsampled through the Patch Merging layer and fused with feature_map_1_2CAF and feature_map_2_3CAF, and then parsed through the third Swin Transformer Block 3 to obtain the feature map feature_map_S3; The feature map feature_map_S3 is downsampled through the Patch Merging layer, and then fused with feature_map_2_3CAF and feature_map_4, and then parsed through the fourth Swin Transformer Block 4 to obtain the feature map feature_map_S4; The feature map feature_map_S4 is fused with feature_map_S3 and the output feature map F1 is subjected to feature extraction and feature fusion by the second RFB module to obtain the feature map feature_map_S3_S4_F; The feature map feature_map_S3_S4_F is fused with feature_map_S2 and the output feature map F1 is subjected to feature extraction and feature fusion by the first RFB module to obtain the feature map feature_map_S2_F; The feature map feature_map_S4 is further subjected to feature extraction and feature fusion by the fourth WCCW module to obtain the feature map feature_map_W4; The feature map feature_map_S3 is connected with feature_map_S3_S4_F and feature_map_W4 in the channel dimension to obtain the feature map G, which is further extracted and fused through the third WCCW module to obtain the feature map feature_map_W3; The feature map feature_map_S2 is connected with feature_map_S2_F and feature_map_W3 in the channel dimension to obtain the feature map G, which is further extracted and fused through the second WCCW module to obtain the feature map feature_map_W2; The feature map feature_map_S1 is connected with feature_map_S2_F and feature_map_W2 in the channel dimension to obtain the feature map G, which is further extracted and fused through the first WCCW module to obtain the coarse feature map feature_map_W1; Input the feature maps feature_map_W1, feature_map_W2, feature_map_W3 and feature_map_W4 into the SSRM Laplacian pyramid module, which will fuse feature information at different levels and output a fine feature map feature_map_all that integrates multi-scale information; The coarse feature map feature_map_W1 and the fine feature map feature_map_all are connected in the channel dimension and input into the detection head to obtain the final result.

3. The remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement according to claim 2, characterized in that: The RFB module is specifically: The feature map F1 is input into the convolution layer of the 1*1 convolution kernel to change the number of channels to reduce the number of parameters and calculations of the model and retain important feature information. Then the output feature map is divided into three branches: F2, F3, and F4. The feature map F2 is input into the convolution layer of the 1*1 convolution kernel to enhance the expressiveness of the features and output the feature map F2_1. The feature map F3 is input into the convolution layer of the 3*3 convolution kernel to capture more complex local features and output the feature map F3_1. The feature map F4 is input into the convolution layer of the 5*5 convolution kernel to capture a larger range of local features and obtain a wider range of contextual information and output the feature map F4_1. Figure F2_1 is input into the dilated convolution layer with a dilation rate of 1 to extract contextual information and output the feature map F2_2; the feature map F3_1 is input into the dilated convolution layer with a dilation rate of 3 to expand the receptive field and extract a wide range of contextual information, and output the feature map F3_2; the feature map F4_1 is input into the dilated convolution layer with a dilation rate of 5 to further expand the receptive field and extract a wider range of contextual information, and output the feature map F4_2; finally, the feature maps F2_2, F3_2, and F4_2 of the three branches are connected in the channel dimension; in this way, the features of different scales and different receptive fields can be fused to obtain a feature representation that integrates multi-scale information.

4. The remote sensing semantic segmentation algorithm based on multi-scale local feature enhancement according to claim 3, characterized in that: The WCCW module is specifically: Input the feature map G into the window partition module to divide the input feature map into multiple local windows and output the feature map G1; pass the feature map G1 through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and output the feature map G1_1; finally, input the feature map G1_1 into the self-attention module to enhance the feature representation of the feature map and capture the global dependency between different regions in the feature map, and output the feature map G1_2; Input the feature map G into the cross window partition module to divide the feature map into local windows and perform cross-window information interaction so that different windows can share information, and output the feature map G2; pass the feature map G2 through the IDsquence module to generate a unique corresponding identifier for the feature map corresponding to each window, and output the feature map G2_1; finally, input the feature map G2_1 into the self-attention module to enhance cross-window information interaction and extract hierarchical features, and output the feature map G2_2; Connect the feature map G1_2 and the feature map G2_2 in the channel dimension to enhance their feature representation. Integrate global and local feature representations and allow the model to simultaneously utilize fine-grained local features and large-scale global context information to improve the detection capability of small objects, and output feature map G_12; The feature map G_12 passes through the convolution layer with a 1*1 convolution kernel, then passes through the Batch Normalization layer, and finally inputs it into the DW convolution layer to further optimize the feature representation, improve the performance and computational efficiency of the model, and output the feature map G_all.

Images