A remote sensing image semantic segmentation method and device based on wavelet transform convolution

Abstract

The application discloses a kind of based on wavelet transform convolution remote sensing image semantic segmentation method and device, the method is first based on wavelet transform convolution encoder extracts semantic feature information in remote sensing image;Second, using double cross attention to connect encoder and decoder, the attention is through the use of channel cross attention and spatial cross attention in series, enhance the feature information of encoder output, enhance the effectiveness of the jump connection between encoder and decoder, reduce their semantic gap;Finally, using traditional CNN decoder for feature map to obtain prediction result, the prediction result is remote sensing image semantic segmentation result.The application can efficiently capture semantic context information in remote sensing image, enhance the response of model to shape feature, improve the precision of semantic segmentation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing image processing, and particularly relates to a remote sensing image semantic segmentation method and device based on wavelet transform convolution. BACKGROUND

[0002] Remote sensing image semantic segmentation is a key task in computer vision and remote sensing, which involves assigning each pixel in a remote sensing image to a predefined class label, such as city, forest, water body, etc. This technology is crucial for environmental monitoring, urban planning, agricultural management, and disaster assessment, etc. Traditional remote sensing image semantic segmentation methods usually rely on manual feature extraction and classic machine learning algorithms. Traditional methods often require expert knowledge and manual feature extraction, which is time-consuming and complex, and difficult to adapt to new data sets or changing environmental conditions. Moreover, they often have poor generalization ability and are difficult to adapt to the rich ground feature characteristics in modern high-precision remote sensing images.

[0003] Currently, the more mainstream method is based on deep learning. Deep learning, especially convolutional neural networks (CNNs), has become the core technology of remote sensing image semantic segmentation. These models can automatically learn complex feature representations from large amounts of data without relying on manual feature extraction, which is an essential step in traditional methods. Deep learning models learn mapping directly from pixel-level labels through end-to-end training, improving the accuracy and efficiency of segmentation. SUMMARY

[0004] In view of the efficiency problem of remote sensing semantic segmentation, the purpose of the present application is to provide a remote sensing image semantic segmentation method and device based on wavelet transform convolution, which captures semantic context information in features by using wavelet transform convolution, enhances the distinguishability of ground feature classification, and achieves the purpose of accurate extraction of remote sensing image semantic segmentation targets.

[0005] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0006] In the first aspect, the present application provides a remote sensing image semantic segmentation method based on wavelet transform convolution, comprising:

[0007] Obtaining a remote sensing image to be predicted;

[0008] Inputting the remote sensing image to be predicted into a pre-constructed remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image; the semantic segmentation result of the remote sensing image refers to target ground feature classification of each pixel in the remote sensing image;

[0009] The construction method of the remote sensing image semantic segmentation prediction model is:

[0010] constructing a remote sensing image semantic segmentation network model; the remote sensing image semantic segmentation network model comprises an encoder, a skip connection and a decoder; the encoder is used to extract a semantic information feature map of an input image, the skip connection is used to enhance the encoder output feature, and the decoder is used to output a prediction result; the prediction result refers to a prediction map which is the same in size as the input image, and the prediction map classifies each pixel in the remote sensing image as a target ground object;

[0011] acquiring a remote sensing image benchmark dataset for label annotation, and dividing the remote sensing image benchmark dataset into a training set and a test set, training the remote sensing image semantic segmentation network model based on the training set, and obtaining a network model with optimal parameters as a remote sensing image semantic segmentation prediction model.

[0012] Preferably, before each iteration training, the images in the training set and the test set are subjected to random cropping, random horizontal sliding, random flipping, rotation and color enhancement operations.

[0013] Preferably, the remote sensing image semantic segmentation network model comprises four encoders, four skip connections and four decoders.

[0014] The encoder extracts the semantic information feature map of the input image based on wavelet change convolution.

[0015] The skip connection enhances the semantic information feature map based on channel cross-attention and spatial cross-attention.

[0016] Preferably, the encoder extracts the semantic information feature map of the input image based on wavelet change convolution, comprising:

[0017] extracting features from the input remote sensing image using convolution to obtain a feature map ;

[0018] using wavelet decomposition on the feature map , setting an initial low-pass component , intermediate features , wherein is an initial weight, four fluxes of different frequencies are obtained by cascading wavelet decomposition: , , which are respectively the semantic information feature maps obtained by the encoders of different layers, and the fluxes of different frequencies are subjected to convolution operation to obtain four intermediate features of different frequencies: ; wherein, refers to wavelet transform, is convolution operation, is a weight.

[0019] Preferably, the skip connection is based on channel cross-attention and spatial cross-attention to enhance the semantic information feature map, including:

[0020] intermediate features of encoder output Perform block processing to obtain , :

[0021] ,

[0022] in, This represents a two-dimensional average pooling operation. Indicates reshaping, This represents a one-dimensional depthwise separable convolution;

[0023] right To perform channel cross-attention processing, first generate The three variables, representing the query, key, and value respectively, are as follows:

[0024] , , ,

[0025] The final feature map is obtained as follows: ,

[0026] in, It is a constant. Indicates channel cross attention. refer to Activation function;

[0027] For feature maps To perform spatial cross-attention processing, first generate The three variables, representing the query, key, and value respectively, are as follows:

[0028] , , ;

[0029] The final enhanced semantic information feature map is obtained as follows: ,in, It is a constant.

[0030] Preferably, the decoder outputs prediction results, including:

[0031] Assume hybrid characteristics ,for , ;

[0032] make ,get Where IWT represents inverse wavelet transform,

[0033] The first layer decoder decodes the obtained up-samples the restores the original image size, and uses convolution to output the final prediction result.

[0034] Preferably, in the training process of the remote sensing image semantic segmentation network model, a cross-entropy loss function is used for loss calculation, and model parameters are updated through back propagation.

[0035] The second aspect of the present application also provides a remote sensing image semantic segmentation device based on wavelet transform convolution, which is used to implement the remote sensing image semantic segmentation method based on wavelet transform convolution.

[0036] A data acquisition module is configured to acquire a remote sensing image to be predicted.

[0037] A remote sensing image semantic segmentation prediction model training module is configured to train a remote sensing image semantic segmentation prediction model.

[0038] A prediction output module is configured to input the remote sensing image to be predicted into the remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image; the semantic segmentation result of the remote sensing image refers to ground object identification and classification for each pixel in the remote sensing image.

[0039] The training process of the remote sensing image semantic segmentation prediction model training module is as follows:

[0040] A remote sensing image semantic segmentation network model is constructed; the remote sensing image semantic segmentation network model comprises an encoder, a skip connection, and a decoder; the encoder is configured to extract a semantic information feature map of an input image, the skip connection is configured to enhance the output feature of the encoder, and the decoder is configured to output a prediction result; the prediction result refers to a prediction map with the same size as the input image, and the prediction map classifies target ground objects for each pixel in the remote sensing image.

[0041] A remote sensing image benchmark data set is acquired for label annotation and is divided into a training set and a test set; the remote sensing image semantic segmentation network model is trained based on the training set to obtain a network model with optimal parameters as a remote sensing image semantic segmentation prediction model.

[0042] In a third aspect, the present application provides a computer-readable storage medium storing one or more programs, the one or more programs including instructions which, when executed by a computing device, cause the computing device to perform any of the remote sensing image semantic segmentation methods based on wavelet transform convolution.

[0043] In a fourth aspect, the present application provides a computing device comprising one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs comprise instructions for performing any one of the methods of wavelet transform convolution-based remote sensing image semantic segmentation method according to the above.

[0044] The present application has the following beneficial effects:

[0045] The present application provides a wavelet transform-based remote sensing image semantic segmentation method, which constructs a remote sensing image semantic segmentation prediction model through an encoder-decoder structure. The remote sensing image semantic segmentation prediction model is combined by a traditional convolutional neural network (CNN) and a wavelet transform. It can obtain a nearly global receptive field without significantly increasing the parameters, thereby efficiently extracting semantic information in the remote sensing image. Secondly, the present application uses double cross-attention to connect the encoder and the decoder. The attention uses channel cross-attention and spatial cross-attention in series to enhance the feature information output by the encoder, enhance the effectiveness of the skip connection between the encoder and the decoder, and reduce the semantic gap between them. Finally, a traditional CNN decoder is used to decode the feature map to obtain the prediction result. The present application can overcome the difficulties caused by the differences between ground object classification targets, efficiently capture semantic context information in the remote sensing image, realize accurate segmentation of the model for ground object classification targets, and improve the accuracy of semantic segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a remote sensing image semantic segmentation prediction model training process schematic diagram in the embodiment of the present application;

[0047] Figure 2 is a remote sensing image semantic segmentation network model structure schematic diagram in the embodiment of the present application;

[0048] Figure 3 is a wavelet transform convolution-based remote sensing image semantic segmentation method flowchart provided by the embodiment of the present application;

[0049] Figure 4 is a wavelet transform convolution encoder data processing process schematic diagram in the embodiment of the present application;

[0050] Figure 5 is a double cross-attention skip connection data processing process schematic diagram in the embodiment of the present application. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions, and advantages of the present application clearer, the following further describes the present application in conjunction with the embodiments and the drawings. Herein, the illustrative embodiments of the present application and their descriptions are used to explain the present application, but not as a limitation of the present application.

[0052] It is also necessary to point out that, in order to avoid obscuring the application with unnecessary details, only the structures and / or processing steps that are closely related to the solution according to the application are shown in the drawings, while other details that are not relevant to the application are omitted.

[0053] It should be emphasized that the term "comprises / comprising" when used in this text refers to the presence of the features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0054] In the following, embodiments of the application will be described with reference to the accompanying drawings. In the drawings, like reference numerals designate identical or similar parts, or identical or similar steps.

[0055] It is emphasized here that the step designations mentioned in the following are not a limitation of the order of the steps, but it is to be understood that the steps can be carried out in the order mentioned in the embodiments, but also differently from the order in the embodiments, or several steps can be carried out simultaneously.

[0056] The first embodiment of the application provides a remote sensing image semantic segmentation method based on wavelet transform convolution, comprising:

[0057] Obtaining a remote sensing image to be predicted;

[0058] Inputting the obtained remote sensing image into a pre-constructed remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image.

[0059] It should be noted that the remote sensing image to be predicted is obtained in the following way:

[0060] The present case is to perform semantic segmentation on high-resolution remote sensing images, such as images in the ISPRS Potsdam and Vaihingen data sets. Such remote sensing images can be obtained on a UAV remote sensing platform and contain target ground objects such as cars, trees, shrubs, buildings and hard road surfaces, as well as other background ground objects.

[0061] It should be noted that the semantic segmentation result of the remote sensing image includes:

[0062] The semantic segmentation of the remote sensing image performs pixel-level image classification on the remote sensing image, and the semantic segmentation result of the remote sensing image is a prediction map with the same size as the remote sensing image, which classifies target ground objects for each pixel in the image. The prediction map labels target ground object pixels, and different colors represent different ground object types. For example, yellow represents cars, green represents trees, blue represents buildings, and red pixels represent the background, etc.

[0063] Reference is made to Figure 1In one embodiment, a remote sensing image semantic segmentation prediction model is constructed in advance, and the specific implementation is as follows:

[0064] S1, a remote sensing image semantic segmentation network model is constructed; see Figure 2 The remote sensing image semantic segmentation network model comprises four encoders, four skip connections and four decoders.

[0065] The encoder is configured to extract a semantic information feature map. In this embodiment, the encoder uses wavelet change convolution. It can make the convolutional neural network obtain a nearly global receptive field without significantly increasing the parameters by combining wavelet transform to expand the receptive field. The input is decomposed into different frequency bands by wavelet decomposition, allowing the convolution layer to process low-frequency and high-frequency components separately, and enhancing the model's response to low-frequency components (i.e. shape features).

[0066] The skip connection is configured to enhance the encoder output feature. In this embodiment, the skip connection uses double cross-attention. With the channel cross-attention (CCA) and spatial cross-attention (SCA) therein, the effectiveness of the U-Net skip connection is enhanced, richer global context information is captured, and the model can better extract fine-grained features.

[0067] The decoder is configured to process the encoder output feature to obtain a prediction result, which is the semantic segmentation result of the remote sensing image. (The prediction result here refers to a prediction map with the same size as the input image, which classifies each pixel in the remote sensing image as a target ground object. The prediction map generated by the prediction module is used to compare and calculate the loss with the real label map to optimize the model in the model training stage; and the semantic segmentation result of the remote sensing image is generated in the model prediction stage).

[0068] S2, a remote sensing image benchmark dataset is obtained for label annotation, and is divided into a training set and a test set. The training set is input into the constructed remote sensing image semantic segmentation network model to extract and enhance the semantic feature information of the input image, and output a prediction result; wherein the label refers to the classification of each pixel in the remote sensing image, and the purpose of using the real label is to compare and calculate the model prediction result to optimize the model in the model training stage; the prediction result refers to a prediction map with the same size as the input image, which classifies each pixel in the remote sensing image as a target ground object;

[0069] S3, based on the loss between the prediction result and the real label, the parameters of the constructed remote sensing image semantic segmentation network model are updated, and the model is iteratively trained until the maximum iteration number is reached or the model reaches the convergence condition, to obtain a network model with optimal parameters as the remote sensing image semantic segmentation prediction model.

[0070] S4, the performance of the constructed remote sensing image semantic segmentation prediction model is tested by using the test set.

[0071] In this embodiment, the ISPRS Potsdam dataset is selected as the benchmark dataset for remote sensing images.

[0072] It should be noted that before each iteration, random cropping, random horizontal sliding, random flipping, rotation, and color enhancement operations are performed on the images in the training and test sets to enhance the model's generalization performance.

[0073] It should be noted that during training, the cross-entropy loss function is used to calculate the difference between the predicted result and the true label, and the model parameters are updated through backpropagation to minimize the loss value. A maximum number of iterations is set; after training reaches the set maximum number of iterations, training stops, and the best-performing model parameters are saved and imported into the network model to obtain the remote sensing image semantic segmentation prediction model. The cross-entropy loss function is shown in the formula:

[0074] ,

[0075] in This is the true label of the sample, with a value of 0 or 1. It represents the predicted probability. The cross-entropy loss function measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels.

[0076] In one embodiment, the maximum number of iterations is set to 200.

[0077] In one embodiment, the asymmetric semantic enhancement module is based on asymmetric convolution and global average pooling operations. Global average pooling can integrate global information into semantic context information, thereby improving segmentation accuracy. Asymmetric convolution can reduce computational cost and improve computational efficiency without changing the receptive field size.

[0078] See Figure 4 The process of extracting semantic information feature maps by wavelet transform convolutional encoder is as follows ( The value can be 1 to 4 (because the model of this invention has 4 layers).

[0079] 11) Let the input remote sensing image be... .in, These represent the height, width, and number of channels of the feature map, respectively. For the preprocessed remote sensing image... use Convolution extracts features to obtain feature maps. ;

[0080] 12) Feature map Using wavelet decomposition, let the initial low-pass component be... intermediate features ,in is a set of 0~1 values randomly generated, and the intermediate feature refers to the feature transferred from the encoder through the skip connection processing to the decoder; for , four fluxes of different frequencies are obtained by cascading wavelet decomposition, (corresponding to the semantic information feature map obtained by different layers in the encoder), and the convolution operation is performed on the fluxes of different frequencies to obtain four intermediate features of different frequencies .

[0081] wherein refers to wavelet transform, and the specific operation is to first give a low-pass filter , a set of high-pass filters , and then take the four filters as weights to perform convolution operation on the input ( ), to obtain four components (among them, is the low-pass component, respectively represent the high-frequency components of the horizontal, vertical and diagonal lines) .

[0082] The advantage of wavelet transform is that it can extract the low-pass component (i.e. shape feature) in the image, and the low-pass component can be enhanced in subsequent processing, thereby strengthening the segmentation ability of the semantic segmentation model.

[0083] Referring to Figure 5 , the process of double cross-attention skip connection to enhance features is as follows (wherein ):

[0084] 21) the intermediate feature output by the encoder is processed by block, i.e. a complete 2D image is decomposed into sequence information, represented as:

[0085] ,

[0086] wherein, represents a two-dimensional average pooling operation, represents reshaping, represents a one-dimensional depth separable convolution.

[0087] 22) the channel cross-attention (CCA) processing is performed on , three variables are first generated, respectively denoted as query, key and value, represented as:

[0088] , , ,

[0089] and finally the feature map is obtained: ,

[0090] wherein, is a constant, CCA represents channel cross attention, denotes an activation function.

[0091] 23) the feature map is processed by spatial cross attention (SCA), first three variables are generated, denoted as query, key, value, respectively, denoted as:

[0092] , , ;

[0093] Finally, the enhanced semantic information feature map is obtained: wherein, is a constant.

[0094] In the embodiment of the application, the decoder processes the value transmitted by the skip connection and makes a prediction, the specific operation being as follows:

[0095] 31) set the mixed feature , for , ; make , get ; wherein IWT represents inverse wavelet transform, which is essentially a transposed convolution, which can convert the frequency domain feature back to the spatial domain feature, because the foregoing wavelet transform (WT) converts the feature from the spatial domain to the frequency domain. In the decoder, the low-frequency feature transmitted by the skip connection in the same layer is added to the mixed feature transmitted by the previous layer, and then the inverse wavelet transform operation is performed together with the high-frequency flux, so that the enhanced feature information can be obtained.

[0096] 32) further, in the first layer decoder, the obtained is upsampled to restore to the original image size, and convolution is used to output the final prediction result. The prediction result is a prediction map with the same size as the remote sensing image, which classifies the target ground objects in each pixel of the image.

[0097] Based on the above inventive concept, a second embodiment of the application provides a remote sensing image semantic segmentation device based on wavelet transform convolution, which is used to realize the remote sensing image semantic segmentation method based on wavelet transform convolution. The device comprises:

[0098] A data acquisition module is configured to acquire a remote sensing image to be predicted. ​

[0099] The remote sensing image semantic segmentation prediction model training module is configured to train a remote sensing image semantic segmentation prediction model, which includes an encoder, a skip connection, and a decoder. The encoder uses wavelet transform convolution, which can expand the receptive field by combining wavelet transform, so that the convolutional neural network can obtain a nearly global receptive field without significantly increasing the parameters. The input is divided into different frequency bands using wavelet decomposition, allowing the convolutional layer to process low-frequency and high-frequency components separately, enhancing the model's response to low-frequency components (i.e., shape features). The skip connection uses double cross-attention, which enhances the effectiveness of the U-Net skip connection by using channel cross-attention (CCA) and spatial cross-attention (SCA) to capture richer global context information, enabling the model to better extract fine-grained features. The decoder is configured to process the decoder output features to obtain a prediction result, which is the semantic segmentation result of the remote sensing image.

[0100] The remote sensing image semantic segmentation prediction model training process is described in detail below. Figure 3 , and the details are as follows:

[0101] S1, a remote sensing image semantic segmentation network model is constructed; see Figure 2 The remote sensing image semantic segmentation network model includes four encoders, four skip connections, and four decoders.

[0102] S2, a remote sensing image benchmark dataset is obtained for label annotation, and is divided into a training set and a test set. The training set is input into the constructed remote sensing image semantic segmentation network model, the semantic feature information of the input image is extracted using wavelet transform convolution-based encoding, the semantic information feature map is obtained using double cross-attention-based skip connection, and the prediction result is output using the convolutional network-based decoder. The label refers to the classification of each pixel in the remote sensing image. The purpose of using the real label is to compare and calculate the model prediction result during the model training phase to optimize the model. The prediction result refers to a prediction map with the same size as the input image, which classifies the target ground objects in the remote sensing image.

[0103] S3, based on the loss between the prediction result and the real label, the parameters of the constructed remote sensing image semantic segmentation network model are updated, and the training is iterated continuously until the maximum iteration number is reached or the model reaches the convergence condition, obtaining the network model with optimal parameters as the remote sensing image semantic segmentation prediction model.

[0104] S4, the performance of the constructed remote sensing image semantic segmentation prediction model is tested using the test set.

[0105] The device further comprises a prediction output module configured to input the obtained remote sensing image to be predicted into a pre-constructed remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image.

[0106] It should be noted that the semantic segmentation result of the remote sensing image refers to that the remote sensing image semantic segmentation performs pixel-level image classification on the remote sensing image, and the semantic segmentation result of the remote sensing image is a prediction map with the same size as the remote sensing image, which classifies target ground objects for each pixel in the image. The prediction map labels target ground object pixels, and different colors represent different ground object types. For example, yellow represents a car, green represents a tree, blue represents a building, and red pixels represent a background.

[0107] The specific working processes of the modules described above can refer to the corresponding processes in the foregoing method embodiments, and will not be described here. The division of the modules is only a logical functional division, and another division mode can be used in actual implementation, for example, a plurality of modules can be combined or integrated into another system.

[0108] Based on the same inventive concept, the present application provides a computer system, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded into the processor, implements the steps of the wavelet transform convolution-based remote sensing image semantic segmentation method described above.

[0109] Based on the same inventive concept, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the wavelet transform convolution-based remote sensing image semantic segmentation method described above.

[0110] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0111] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flowsheet block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0112] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flowsheet block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0113] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flowsheet block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0114] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing the technical solutions of the present application, but not for limiting it. Although the present application is described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement should be covered in the protection scope of the claims of the present application.

Claims

1. A method for remote sensing image semantic segmentation based on wavelet transform convolution, characterized in that, The method comprises the steps of: acquiring a remote sensing image to be predicted; inputting the remote sensing image to be predicted into a pre-constructed remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image; the semantic segmentation result of the remote sensing image refers to target ground object classification of each pixel in the remote sensing image; the remote sensing image semantic segmentation prediction model is constructed in the following manner: a remote sensing image semantic segmentation network model is constructed; the remote sensing image semantic segmentation network model comprises four encoders, four skip connections and four decoders; the encoder is used to extract a semantic information feature map of an input image, the skip connection is used to enhance an encoder output feature, and the decoder is used to output a prediction result; the prediction result refers to a prediction image with the same size as the input image, and the prediction image classifies target ground objects in each pixel in the remote sensing image; The encoder extracts a semantic information feature map of an input image based on wavelet transform convolution, including: performing wavelet transform on an input remote sensing image Using Convolution extracts features to obtain a feature map ; To the feature map Using wavelet decomposition, set the initial low-pass component , intermediate features , wherein The initial weight is obtained by cascading wavelet decomposition to obtain four fluxes of different frequencies: , , respectively, the semantic information feature map obtained by the encoder of different layers, and the convolution operation is performed on the fluxes of different frequencies to obtain four intermediate features of different frequencies: ; wherein, Indicates wavelet transform, Convolution operation, Weight; the skip connection is based on channel cross-attention and spatial cross-attention to enhance the semantic information feature map; a remote sensing image benchmark dataset is acquired to perform label annotation and is divided into a training set and a test set; the remote sensing image semantic segmentation network model is trained based on the training set to obtain a network model with optimal parameters as the remote sensing image semantic segmentation prediction model. 2.The method of claim 1, wherein, Before each iteration training, the images in the training set and the test set are subjected to random cropping, random horizontal sliding, random flipping, rotation and color enhancement operations. 3.The method of claim 1, wherein, the skip connection is based on channel cross-attention and spatial cross-attention to enhance the semantic information feature map, which comprises: Intermediate features of the encoder output Chunking is performed to obtain , : ， wherein, denotes a two-dimensional average pooling operation, denotes reshaping, denotes a one-dimensional depthwise separable convolution; To perform channel cross attention processing, first generate three variables, respectively denoted as query, key, value, denoted as: ， ， ， The final feature map is obtained: , wherein, is a constant, denotes channel cross attention, refers to activation function; To the feature map Spatial cross attention processing is performed, first generating Three variables, denoted as query, key, and value, are represented as: ， ， ； The enhanced semantic information feature map is finally obtained: wherein, is a constant.

4. The wavelet transform convolution-based remote sensing image semantic segmentation method according to claim 3, characterized in that, the decoder outputs the prediction result, which comprises: Set mixed features For , ; causing , obtaining ; wherein IWT denotes inverse wavelet transform, The first layer decoder upsamples the obtained to the original image size and uses convolution to output the final prediction.​ 5. The wavelet transform convolution-based remote sensing image semantic segmentation method according to claim 1, characterized in that, during the training process of the remote sensing image semantic segmentation network model, a cross-entropy loss function is used for loss calculation, and the model parameters are updated in a back propagation manner.

6. A device for semantic segmentation of a remote sensing image based on wavelet transform convolution, characterized in that, The device for implementing the remote sensing image semantic segmentation method based on wavelet transform convolution according to any one of claims 1 to 5 comprises: a data acquisition module configured to acquire a remote sensing image to be predicted; a remote sensing image semantic segmentation prediction model training module configured to train a remote sensing image semantic segmentation prediction model; a prediction output module configured to input the remote sensing image to be predicted into the remote sensing image semantic segmentation prediction model to obtain a semantic segmentation result of the remote sensing image; the remote sensing semantic segmentation result refers to ground object identification and classification of each pixel in the remote sensing image; the training process of the remote sensing image semantic segmentation prediction model training module is as follows: a remote sensing image semantic segmentation network model is constructed; the remote sensing image semantic segmentation network model comprises four encoders, four skip connections and four decoders; the encoder is used to extract a semantic information feature map of an input image, the skip connection is used to enhance an encoder output feature, and the decoder is used to output a prediction result; the prediction result refers to a prediction image with the same size as the input image, and the prediction image classifies target ground objects in each pixel in the remote sensing image; The encoder extracts a semantic information feature map of an input image based on wavelet transform convolution, including: performing wavelet transform on an input remote sensing image Using Convolution extracts features to obtain a feature map ; To the feature map Using wavelet decomposition, set the initial low-pass component , intermediate features , wherein The initial weight is obtained by cascading wavelet decomposition to obtain four fluxes of different frequencies: , , respectively, the semantic information feature map obtained by the encoder of different layers, and the convolution operation is performed on the fluxes of different frequencies to obtain four intermediate features of different frequencies: ; wherein, Indicates wavelet transform, Convolution operation, Weight; the skip connection is based on channel cross-attention and spatial cross-attention to enhance the semantic information feature map; A remote sensing image benchmark data set is acquired for label annotation, and is divided into a training set and a test set. The remote sensing image semantic segmentation network model is trained based on the training set, and a network model with optimal parameters is obtained as a remote sensing image semantic segmentation prediction model.

7. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions that when executed by a computer cause the computer to perform a method of any of claims 1-6. The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the wavelet transform convolution-based remote sensing image semantic segmentation methods of claims 1-5.

8. A computing device, comprising: Comprise, One or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the wavelet transform convolution-based remote sensing image semantic segmentation methods of claims 1-5.

Images