Remote sensing image semantic segmentation method and system based on ConvNeXt convolution
By building a U-Net network based on ConvNeXt and combining parallel feature extraction and attention feature fusion, the problems of high resolution and data imbalance in remote sensing image segmentation are solved, and high-precision semantic segmentation and improved computational efficiency are achieved.
Patent Information
- Application Number
- CN202310491829.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-26
- Publication Date
- 2025-09-09
- Estimated Expiration
- 2043-04-26
AI Technical Summary
Semantic segmentation algorithms for remote sensing images are difficult to effectively process high-resolution images. Traditional convolutional models are computationally intensive and difficult to achieve satisfactory results. Unbalanced data distribution leads to low segmentation accuracy.
The ConvNeXt module is used to construct the U-Net network, combined with parallel feature extraction and attention feature fusion modules, global features are extracted through large convolution kernels, local features are extracted using the ResNet module, and information loss is reduced through skip connections. The DConv and UConv modules are designed to reduce the amount of computation.
It improves the accuracy of remote sensing image segmentation, reduces the amount of calculation, solves the problem of uneven data distribution, and enhances the robustness of the model and the segmentation effect.
Smart Images
Figure CN116580195B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision, and in particular relates to a remote sensing image semantic segmentation method and system based on ConvNeXt convolution. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] Semantic segmentation of remote sensing images is an important tool for processing remote sensing images. With the development of aerial photography, aerial scanning, and microwave radar technologies, the demand for remote sensing image processing is increasing. Semantic segmentation of remote sensing images can better serve land cover monitoring, for example, wetland resource monitoring by extracting wetland boundaries for wetland classification; forest cover monitoring by assessing changes in forest resources and ecological status; and grassland cover monitoring by quantitatively classifying grasslands.
[0004] Remote sensing images have many targets and they are densely arranged. The sizes of similar targets vary greatly, and the color and texture are quite different, which makes it difficult for general segmentation algorithms to achieve satisfactory results. At the same time, due to the continuous improvement of their spectral resolution, the computational complexity of the segmentation model has also greatly increased. It is difficult for traditional convolution models to achieve good results, and improvements are needed in the model structure. Summary of the Invention
[0005] To overcome the shortcomings of the above-mentioned prior art, the present invention provides a remote sensing image semantic segmentation method and system based on ConvNeXt convolution, constructs a U-shaped network structure (ConvNeXt Embedding U-NetNetwork, CE-UNet) based on the ConvNext module, which is used to perform semantic segmentation on remote sensing images and solve the problems of large resolution and unbalanced data distribution of remote sensing images.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:
[0007] The first aspect of the present invention provides a remote sensing image semantic segmentation method based on ConvNeXt convolution;
[0008] The semantic segmentation method of remote sensing images based on ConvNeXt convolution includes:
[0009] Acquire remote sensing images and perform expansion processing to obtain training data sets;
[0010] Based on the training dataset, the semantic segmentation model built based on ConvNeXt is trained;
[0011] Obtain the remote sensing image to be segmented, perform semantic segmentation using the trained semantic segmentation model, and obtain the segmentation result;
[0012] Among them, the semantic segmentation model obtains the fused features based on the feature extraction and fusion module. The feature extraction and fusion module includes a parallel feature extraction module and an attention feature fusion module. The parallel feature extraction module is composed of a ConvNext module and a ResNet module, and performs parallel extraction of global features and local features; the attention feature fusion module uses the attention mechanism to fuse global features and local features.
[0013] Furthermore, the expansion process is specifically: randomly scaling and randomly cropping the remote sensing image, and then performing rotation and normalization operations to generate more training samples to form a training data set.
[0014] Furthermore, the overall architecture of the semantic segmentation model is a U-Net architecture, with downsampling performed on the left and upsampling performed on the right; skip connections are used between downsampling and upsampling, and images of the same size are connected through skip connections to compensate for information loss.
[0015] Furthermore, the semantic segmentation model further includes a DConv module and a UConv module;
[0016] The DConv module uses an even number of convolution kernels to downsample and reduce the resolution of the image;
[0017] The UConv module uses an even-numbered convolution kernel to perform upsampling to restore the image resolution.
[0018] Furthermore, in the parallel feature extraction module, the input image is first cropped in terms of the number of channels through a Split operation, half of the image with the number of channels is subjected to a ConvNeXt module to extract global features, and the other half of the image with the number of channels is subjected to a ResNet module to extract local features.
[0019] Furthermore, the specific operations of the ConvNeXt module are:
[0020] The input image is subjected to a large convolution, followed by normalization and a 1×1 convolution. The image is then processed using a GELU activation function, followed by a 1×1 convolution for dimensionality reduction. Finally, the SkipConnected operation is used to add the input image and the image after dimensionality reduction to obtain global features.
[0021] Furthermore, the pytorch_lighting library is used to train the semantic segmentation model to obtain the optimal parameters of the model.
[0022] The second aspect of the present invention provides a remote sensing image semantic segmentation system based on ConvNeXt convolution.
[0023] The remote sensing image semantic segmentation system based on ConvNeXt convolution includes data expansion module, model training module and semantic segmentation module:
[0024] The data expansion module is configured to: acquire remote sensing images and perform expansion processing to obtain a training data set;
[0025] The model training module is configured to: train the semantic segmentation model built based on ConvNeXt based on the training dataset;
[0026] The semantic segmentation module is configured to: obtain the remote sensing image to be segmented, perform semantic segmentation using the trained semantic segmentation model, and obtain the segmentation result;
[0027] Among them, the semantic segmentation model obtains the fused features based on the feature extraction and fusion module. The feature extraction and fusion module includes a parallel feature extraction module and an attention feature fusion module. The parallel feature extraction module is composed of a ConvNext module and a ResNet module, and performs parallel extraction of global features and local features; the attention feature fusion module uses the attention mechanism to fuse global features and local features.
[0028] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the remote sensing image semantic segmentation method based on ConvNeXt convolution as described in the first aspect of the present invention.
[0029] The fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the remote sensing image semantic segmentation method based on ConvNeXt convolution as described in the first aspect of the present invention are implemented.
[0030] One or more of the above technical solutions have the following beneficial effects:
[0031] The present invention provides a remote sensing image semantic segmentation method and system based on ConvNeXt convolution, constructing a U-shaped network structure (ConvNeXt Embedding U-Net Network, CE-UNet) based on the ConvNext module, which is used to perform semantic segmentation on remote sensing images and solve the problems of large remote sensing image resolution and unbalanced data distribution.
[0032] In this invention, the ConvNeXt module is used to replace the Transformer for global feature sampling. In view of the problem that the computational complexity of the Transformer's attention mechanism increases exponentially with the increase of image size, the use of the ConvNeXt module can not only play a replacement role, but also significantly reduce the computational complexity.
[0033] In this paper, a parallel feature extraction structure is designed. By performing a split operation on the number of channels and utilizing the local feature extraction capability of the ResNet module and the global feature extraction capability of the ConvNeXt module, the network can obtain global and local information at the same time. The proposed AFM module is then used to fuse the feature information, effectively improving the segmentation accuracy of the model.
[0034] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0036] Figure 1 This is a flow chart of the method of the first embodiment.
[0037] Figure 2 This is a structural diagram of the semantic segmentation model of the first embodiment.
[0038] Figure 3 This is the structural diagram of the feature extraction and fusion module of the first embodiment.
[0039] Figure 4 This is a structural diagram of the ConvNeXt module of the first embodiment.
[0040] Figure 5 This is a structural diagram of the ResNet module of the first embodiment.
[0041] Figure 6 This is the structural diagram of the attention feature fusion module of the first embodiment. DETAILED DESCRIPTION
[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0043] Traditional CNN networks mainly process images through the local information extraction capability of convolution to achieve classification, detection and segmentation. Due to the large image resolution of remote sensing images, the effect of only extracting local information for segmentation is poor. The emergence of ViT officially introduced the Transformer into the field of vision, and it has become an important method for extracting global information from images. However, the Transformer-based network needs to learn from a large number of training sets, and the training sets of remote sensing images themselves are not easy to obtain, which is obviously not very friendly to the training network. So, can we consider improving the traditional convolution module to give it good global information extraction capabilities? When various Transformer networks dominated the charts, ConvNext gave a positive answer to the above question. Using large convolution kernels can also achieve global information extraction capabilities that are not inferior to Transformer, while significantly reducing the amount of computation and the requirement for massive training data.
[0044] With this idea in mind, the present invention designs a parallel feature extraction module (PFE) to extract global information and local information at the same time, so that the network can learn global-local features, combining the strengths of both, making the network's learning effect even better; in order to further enhance the network's information processing capabilities, the present invention designs an attention feature fusion module (AFM) to fuse local information and global information by using the attention mechanism, so that the features are fully integrated, achieving a new effect of successfully extracting occluded objects, such as cars blocked by trees, and improving the model accuracy; at the same time, a DConv downsampling module and a UConv upsampling module are designed, which greatly reduce the amount of computation and prevent overfitting.
[0045] Example 1
[0046] This embodiment discloses a remote sensing image semantic segmentation method based on ConvNeXt convolution;
[0047] like Figure 1 As shown in the figure, the remote sensing image semantic segmentation method based on ConvNeXt convolution includes:
[0048] Step S1: Acquire remote sensing images and perform expansion processing to obtain a training data set.
[0049] Each pixel of the acquired remote sensing image is manually labeled to obtain a labeled dataset.
[0050] Since the resolution and size of remote sensing images are too high, the data needs to be cropped and the images need to be cropped into the same size. Since the data sets of remote sensing images are generally small, in order to obtain more pictures for model learning, it is necessary to use random scaling and random cropping to expand the remote sensing images.
[0051] Specifically, the image is scaled by randomly selecting a value in [0.5, 0.75, 1.0, 1.25, 1.5], and then the image is randomly cropped to 255×255. After unifying the image size, the image is rotated with a probability of 0.5, rotated 90°, and then normalized to generate more training samples, solving the problem that training sets of remote sensing images are difficult to obtain.
[0052] The expanded data constitutes the data training set, and is divided into training set, validation set and test set according to the preset ratio. The training set is used to train the model parameters, the validation set is used to debug the parameters, and the test set is used to test the accuracy of the model. There is no overlap between the three data sets, so the recognition effect of the model can be objectively evaluated.
[0053] Step S2: Based on the training dataset, the semantic segmentation model built based on ConvNeXt is trained, verified and tested.
[0054] The semantic segmentation model is a U-Net network structure (ConvNeXtEmbedding U-Net Network, CE-UNet) based on the ConvNext module; U-Net consists of an encoder and a decoder; the encoder on the left performs downsampling, and the decoder on the right performs upsampling; first, the original image (Raw Image, X r ) is passed into the neural network for overall feature extraction and changing the number of channels of the image, and then passed into the Encoder-Decoder structure; the Encoder is used to extract features, and the Decoder restores the image resolution as finely as possible while fusing high-level semantics and low-level spatial information; because upsampling cannot completely restore the original image, it will cause information loss. In order to reduce information loss, U-Net uses skip connections to connect images of the same size, and uses the decoder to learn the spatial correlation of the corresponding encoding stage through skip connections (Skip Connection), thereby retaining the image after the features are collected at each stage, allowing the network to learn the fusion features of different stages and make up for the information loss. The U-Net network can effectively solve problems such as uneven data distribution and difficulty in extracting image detail information. It has high acquisition capabilities and strong robustness.
[0055] Figure 2 It is the structural diagram of the semantic segmentation model, such as Figure 2 As shown in the figure, the encoder on the left follows the typical architecture of the convolutional network, including a 3×3 convolution, repeated use of four feature extraction and fusion modules and a DConv module three times, and doubling the number of feature channels in each downsampling step; while the decoder on the right includes upsampling of the feature map in each step, and its structure includes repeated use of four feature extraction and fusion modules and a UConv module three times, four feature extraction and fusion modules, a 3x3 convolution, and a 1×1 convolution.
[0056] Based on the above structure, the processing flow of the semantic segmentation model is as follows:
[0057] The original image input (Raw Image, X r ) is of size C×H×W, where C is the number of channels, H is the height of the image, and W is the width of the image.
[0058] In the first step, a 3×3 convolution is performed to increase the number of channels of the original image, thereby increasing the nonlinearity of the image and improving the expressive power of the model.
[0059] In the second step, four feature extraction and fusion modules are continuously passed in to double the number of feature channels. After that, the result is passed to the DConv module, and the image is downsampled using an even-numbered convolution kernel of 2×2 size to reduce the image resolution. This step is then repeated twice.
[0060] In the third step, the output feature image is passed through four feature extraction and fusion modules, and the result is passed to the UConv module. An upsampling operation is performed using an even-numbered convolution kernel of 2×2 size to restore the image resolution, thereby further reducing the amount of calculation while improving the stability of the network. This step is then repeated twice.
[0061] In the fourth step, the output feature image is passed through four feature extraction and fusion modules, and then the result is passed into a 3×3 convolution to restore the image with the same resolution and number of channels as the original image.
[0062] Finally, a 1×1 convolution is passed in as the segmentation head to segment the target image.
[0063] Figure 3 This is the structural diagram of the feature extraction and fusion module, such as Figure 3As shown in the figure, the feature extraction and fusion module mainly consists of a parallel feature extraction module (PFE) composed of a ConvNext module and a ResNet module, and an attention feature fusion module (AFM). The specific processing flow is as follows:
[0064] First, a 1×1 convolution operation is performed on the incoming feature image to change the number of channels.
[0065] Then, using the split operation, the feature image with the changed number of channels is converted into two 0.5×C×H×W images, which are respectively passed into the ConvNeXt module and the ResNet module for global feature extraction and local feature extraction. The size of the output image feature map does not change, and the size of the input feature map is consistent with the size of the output feature map, both of which are 0.5×C×H×W.
[0066] Afterwards, the output global features and local features are passed to the AFM module for fusion processing. The attention mechanism is used to fully fuse the two feature maps. Compared with the simple concat operation, the effect is greatly improved, and the global information and local information are fused to the greatest extent.
[0067] Finally, the image after the AFM module is added to the feature image after the first 1×1 convolution of the feature extraction and fusion module, and then a 1×1 convolution operation is performed to change the number of channels and output the feature image.
[0068] The parallel feature extraction module uses the Split operation to crop the number of channels, dividing the number of channels of the image into two, and passing them into the ConvNeXt module and ResNet module respectively. Among them, the ConvNeXt module uses a large convolution kernel to collect the global information of the image, and the ResNet module uses a small convolution kernel to collect the local information of the image, thereby making the classification results more accurate; at the same time, using the ResNet module and the ConvNeXt module in parallel has greatly improved the effect of feature extraction on the image compared to using a single ResNet network or ConvNeXt network.
[0069] Figure 4 This is the structural diagram of the ConvNeXt module, such as Figure 4As shown in the figure, the ConvNeXt module is designed in comparison with the SwinTransformer. The image is passed through a large 7×7 convolution kernel, which is compared with the 7×7 window of the Swin Transformer. Since the activation function of the Swin Transformer uses the GELU activation function, ConvNeXt also uses the GELU function for comparison. Under the same circumstances, ConvNeXt can greatly reduce the amount of calculation; after the large convolution and normalization operation, and then undergoing a 1×1 convolution, the size of the output image becomes 2×C×H×W; then the image is processed using a GELU activation function, and then a 1×1 convolution is used for dimensionality reduction so that the input image size and output are consistent, and then the Skip Connected operation is used to add the input and processed images to reduce the information loss caused by convolution. The size of the output image of the ConvNeXt module is C×H×W.
[0070] Figure 5 This is the structural diagram of the ConvNeXt module, such as Figure 5 As shown in the figure, the ResNet module uses small convolution operations to facilitate the collection of local information in the image. First, a 1×1 convolution is used to reduce the number of channels, then a 3×3 convolution is used to extract feature information, and then a 1×1 convolution is used to restore the number of channels. Finally, skip connections are added to reduce the information loss caused by the convolution operation. This allows the model to simultaneously obtain local and global information, enhancing the network's ability to capture information.
[0071] The Attention Fusion Module (AFM) first performs a simple channel-wise concatenation of local and global features, followed by feature extraction using convolution. Batch normalization is then performed to prevent computationally intensive network performance instability, followed by an activation function. The processed image is then fed into three separate paths, two of which undergo downsampling, followed by channel restoration. After reshaping, the attention mechanism is applied to extract global and local features that fully consider pixel correlations, resulting in better feature fusion and improved semantic segmentation accuracy.
[0072] Figure 6 This is the structural diagram of the attention feature fusion module, such as Figure 6 As shown, specifically, first perform a simple concat operation on the two input features, and then perform a ConvBNReLU operation. The specific formula is:
[0073] ConvBNReLU=ReLU(BN(Conv(x)))
[0074] Among them, x represents the feature map after concat splicing. After the above operation, it is divided into three paths. One is the convolution Conv operation to generate x q The other two are first processed by Adaptive Average Pooling and then reshaped, and finally processed by Linear layer to obtain x k 、x v Finally, the following operations are performed to obtain the Output:
[0075] Output=SOFTMAX(x q ×x k )×x v
[0076] The size of the output image is C×H×W, which fully integrates global features and local features to improve accuracy.
[0077] The DConv module and UConv module respectively use even-numbered convolution kernels to reduce the amount of computation and avoid problems such as gradient vanishing or gradient exploding caused by too deep a network layer.
[0078] The image after passing through the encoder and decoder is passed through a semantic segmentation head, that is, a 1×1 convolution, to turn the image into a semantic segmentation image.
[0079] Ablation experiments
[0080] The effectiveness of the model constructed by the present invention is demonstrated through ablation experiments, and the experimental data uses the Vaihingen dataset.
[0081] Vaihingen: The Vaihingen dataset consists of 33 very fine TOP image patches with a spatial average resolution of 24.94×2064 pixels. Each TOP image patch has three multispectral bands (near infrared, red, and green) as well as a digital surface model (DSM) and a surface model (NDSM) with a ground sampling distance (GSD) of 9 cm. This dataset involves five foreground classes (impervious surfaces, buildings, low vegetation, trees, and cars) and one background class (clutter). In the experiments, only TOP image patches were used, without DSM and NDSM. Images with IDs 2, 4, 6, 8, 10, 12, 14, 16, 20, 22, 24, 27, 29, 31, 33, 35, and 38 were used for testing, ID 30 was used for validation, and the remaining 15 were used for training.
[0082] Analysis of ablation experiment results:
[0083] In this experiment, only two control experiments were conducted. One group was used as the basic framework without the AFM module, and the other group used the AFM module to further perform feature fusion to improve accuracy. This experiment used three evaluation indicators, namely F1, mIoU and OA. The calculation methods of the three are as follows:
[0084]
[0085]
[0086]
[0087]
[0088]
[0089] Among them, TP k FP k TN k and FN k They represent true positive, false positive, true negative, and false negative, respectively, for a specific object indexed as class k. They are calculated for all categories including the background class. The ablation experiment results are shown in Table 1.
[0090] Table 1 Ablation experiment results
[0091] Method F1 mIoU OA Baseline 89.07％ 80.51％ 89.45％ Baseline+AFM 90.06％ 82.12％ 90.59％
[0092] By analyzing the ablation experiment results, it was found that after adding the attention feature fusion module, OA, mIoU and F1 were improved to varying degrees. Moreover, after adding the attention feature fusion module, the accuracy of F1 increased by 1.01%, mIoU increased by 1.61%, and OA increased by 1.14%. The above experiments show that the model constructed by the present invention has a great contribution to the improvement of the final accuracy, and the above improvements are all effective.
[0093] Example 2
[0094] This embodiment discloses a remote sensing image semantic segmentation system based on ConvNeXt convolution;
[0095] The remote sensing image semantic segmentation system based on ConvNeXt convolution includes data expansion module, model training module and semantic segmentation module:
[0096] The data expansion module is configured to: acquire remote sensing images and perform expansion processing to obtain a training data set;
[0097] The model training module is configured to: train the semantic segmentation model built based on ConvNeXt based on the training dataset;
[0098] The semantic segmentation module is configured to: obtain the remote sensing image to be segmented, perform semantic segmentation using the trained semantic segmentation model, and obtain the segmentation result;
[0099] Among them, the semantic segmentation model obtains the fused features based on a feature extraction fusion module composed of a parallel feature extraction module and an attention feature fusion module. The parallel feature extraction module is composed of a ConvNext module and a ResNet module, and performs parallel extraction of global features and local features; the attention feature fusion module uses an attention mechanism to fuse global features and local features.
[0100] Example 3
[0101] The purpose of this embodiment is to provide a computer-readable storage medium.
[0102] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the remote sensing image semantic segmentation method based on ConvNeXt convolution as described in the first embodiment of the present disclosure.
[0103] Example 4
[0104] The purpose of this embodiment is to provide an electronic device.
[0105] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the remote sensing image semantic segmentation method based on ConvNeXt convolution as described in the first embodiment of the present disclosure are implemented.
[0106] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.
Claims
1. A remote sensing image semantic segmentation method based on ConvNeXt convolution, characterized by: include: Acquire remote sensing images and perform expansion processing to obtain training data sets; Based on the training dataset, the semantic segmentation model built based on ConvNeXt is trained; Obtain the remote sensing image to be segmented, perform semantic segmentation using the trained semantic segmentation model, and obtain the segmentation result; The semantic segmentation model obtains fused features based on a feature extraction and fusion module. The feature extraction and fusion module includes a parallel feature extraction module and an attention feature fusion module. The parallel feature extraction module is composed of a ConvNeXt module and a ResNet module, and performs parallel extraction of global features and local features. The attention feature fusion module uses an attention mechanism to fuse global features and local features. The overall architecture of the semantic segmentation model is a U-Net architecture, with downsampling on the left and upsampling on the right. Skip connections are used between downsampling and upsampling to connect images of the same size to compensate for information loss. The left encoder consists of a Convolution, repeated use of four feature extraction fusion modules and a combination of DConv modules three times; the decoder on the right includes repeated use of four feature extraction fusion modules and a combination of UConv modules three times, four feature extraction fusion modules, a 3x3 convolution, a Convolution; In the parallel feature extraction module, the input image is first cropped in terms of the number of channels through a Split operation, half of the channels of the image are subjected to a ConvNeXt module to extract global features, and the other half of the channels of the image are subjected to a ResNet module to extract local features.
2. The remote sensing image semantic segmentation method based on ConvNeXt convolution according to claim 1, characterized in that The expansion process specifically includes: randomly scaling and randomly cropping the remote sensing image, and then performing rotation and standardization operations to generate more training samples to form a training data set.
3. The remote sensing image semantic segmentation method based on ConvNeXt convolution according to claim 1, characterized in that: The semantic segmentation model also includes a DConv module and a UConv module; The DConv module uses an even number of convolution kernels to downsample and reduce the resolution of the image; The UConv module uses an even-numbered convolution kernel to perform upsampling to restore the image resolution.
4. The remote sensing image semantic segmentation method based on ConvNeXt convolution according to claim 1, characterized in that The specific operations of the ConvNeXt module are: The input image is subjected to a large convolution, followed by normalization and a 1×1 convolution. The image is then processed using a GELU activation function, followed by a 1×1 convolution for dimensionality reduction. Finally, the SkipConnected operation is used to add the input image and the image after dimensionality reduction to obtain global features.
5. The remote sensing image semantic segmentation method based on ConvNeXt convolution according to claim 1, characterized in that: The pytorch_lighting library is used to train the semantic segmentation model to obtain the optimal parameters of the model.
6. A remote sensing image semantic segmentation system based on ConvNeXt convolution, characterized by: Including data expansion module, model training module and semantic segmentation module: The data expansion module is configured to: acquire remote sensing images and perform expansion processing to obtain a training data set; The model training module is configured to: train a semantic segmentation model built based on ConvNeXt based on a training dataset; The semantic segmentation module is configured to: obtain a remote sensing image to be segmented, perform semantic segmentation using a trained semantic segmentation model, and obtain a segmentation result; The semantic segmentation model obtains fused features based on a feature extraction and fusion module. The feature extraction and fusion module includes a parallel feature extraction module and an attention feature fusion module. The parallel feature extraction module is composed of a ConvNeXt module and a ResNet module, and performs parallel extraction of global features and local features. The attention feature fusion module uses an attention mechanism to fuse global features and local features. The overall architecture of the semantic segmentation model is a U-Net architecture, with downsampling on the left and upsampling on the right. Skip connections are used between downsampling and upsampling to connect images of the same size to compensate for information loss. The left encoder consists of a Convolution, repeated use of four feature extraction fusion modules and a combination of DConv modules three times; the decoder on the right includes repeated use of four feature extraction fusion modules and a combination of UConv modules three times, four feature extraction fusion modules, a 3x3 convolution, a Convolution; In the parallel feature extraction module, the input image is first cropped in terms of the number of channels through a Split operation, half of the channels of the image are subjected to a ConvNeXt module to extract global features, and the other half of the channels of the image are subjected to a ResNet module to extract local features.
7. An electronic device, comprising: a memory for non-transitory storage of computer-readable instructions; as well as a processor for executing said computer-readable instructions, When the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 5 is executed.
8. A storage medium, characterized in that: Computer-readable instructions are non-transitory stored, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1 to 5 are executed.