Remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism

Abstract

The present invention discloses a remote sensing image fusion method with multi-scale feature enhancement based on a large-core attention mechanism, wherein the local and non-local feature extraction module is composed of a multi-scale residual block and a large-core attention module, and a pyramid squeeze attention module is used to fully integrate the local and non-local information of the spatial and channel dimensions extracted by the large-core attention module and the multi-scale residual block, and then a U-shaped feature extraction network based on asymmetric convolution is used to obtain low-frequency features of different scales and depths, enhance the semantics of features in the encoding and decoding process, and fully integrate context information. The present invention extracts features of remote sensing images from channel and spatial perspectives by combining local and non-local methods, and constructs a pyramid squeeze attention module to fully integrate the extracted features, and uses ACUNet to enhance the semantic information of features and fully integrate context, thereby greatly improving the image quality of the multispectral and panchromatic image fusion results.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image fusion method based on a large kernel attention mechanism and multi-scale feature enhancement. Background Art

[0002] Remote sensing images have very important application value in the military and national economy. However, due to the limitations of sensor imaging technology, it is impossible to obtain remote sensing images with rich spatial resolution and spectral resolution at the same time. Therefore, satellites are usually equipped with a variety of different sensors to obtain images with different characteristics, and then use image fusion technology to combine multiple images with complementary features to obtain the required high-quality images. Multispectral and panchromatic image fusion technology has always been a hot research topic in the field of remote sensing. Its main purpose is to fuse multispectral images with high spectral resolution and panchromatic images with high spatial resolution in the same scene to obtain multispectral images with high spatial resolution. The results generated by the multispectral and panchromatic image fusion algorithm are also widely used in various aspects of regional science, becoming an effective technical means in the fields of geological mapping, agricultural remote sensing, environmental monitoring, etc.

[0003] In recent years, with the successful application of deep learning in the field of computer vision, its network has shown great potential in feature extraction and nonlinear fitting. Therefore, a large number of multispectral and panchromatic remote sensing image fusion algorithms based on deep learning have been proposed. Compared with traditional remote sensing image fusion algorithms, deep learning-based methods alleviate the shortcomings of traditional image algorithms, such as slow testing time, low adaptability of data sets, and the generation of serious spatial and spectral distortions, which greatly improves the fusion results in terms of spectral preservation and spatial enhancement. Remote sensing image fusion methods based on deep learning are a major innovation in remote sensing image fusion. These innovative algorithms are based on the characteristics of multispectral images and panchromatic images, design appropriate deep learning frameworks, and generate better image results. When performing remote sensing image fusion, spectral consistency and spatial preservation are important indicators for measuring the algorithm fusion performance. How to better balance the two is the most important direction of remote sensing image fusion algorithm research. With the continuous development of deep learning, the quality of fusion results has gradually improved. Therefore, remote sensing image fusion algorithms based on deep learning have attracted the attention of many researchers.

[0004] Recently, due to the amazing performance of transformers in the field of computer vision, people have gradually realized the importance of long-range information. Zhou et al. first applied transformers to the field of remote sensing image fusion and designed an effective image fusion algorithm. They used a specific transformer to extract the corresponding long-range dependencies and used a reversible neural network to complete the fusion of local and non-local features. GR et al. proposed a transformer-based self-attention architecture (Pansformer) and achieved good fusion results. Su et al. proposed a DR-NET model that uses Swintransformer to extract long-range features and uses an attention mechanism to focus the network on spatial and spectral information.

[0005] Although transformers have obvious advantages in extracting long-range dependencies, these transformer-based fusion algorithms still have certain limitations: 1) The cost of quadratic computational complexity is too high for high-resolution remote sensing images. 2) For high-dimensional spectral images with multiple bands, transformers only capture spatial adaptability and ignore channel adaptability. 3) For local information and non-local information, only simple splicing and fusion are performed, ignoring the differences between different information. Summary of the invention

[0006] The present invention provides a remote sensing image fusion method with multi-scale feature enhancement based on a large kernel attention mechanism, which solves the defects of the existing remote sensing image fusion method. It extracts features of remote sensing images from channel and spatial perspectives by combining local and non-local methods, and constructs a pyramid squeeze attention module to fully integrate the extracted features. It uses ACUNet to enhance the semantic information of the features and fully fuse the context, thereby greatly improving the image quality of the multispectral and panchromatic image fusion results.

[0007] The present invention proposes a remote sensing image fusion method based on a large kernel attention mechanism for multi-scale feature enhancement, the method comprising:

[0008] S1: Preprocess the low-resolution multispectral image and panchromatic image and perform preliminary feature extraction through two simple convolution blocks;

[0009] S2: construct a local and non-local feature extraction module consisting of a multi-scale residual block and a large core attention module. The large core attention module is mainly used to obtain long-range dependencies, local spatial information, and adaptability of spatial and spectral dimensions. The multi-scale residual block is mainly used to extract multi-scale local spatial information and perform preliminary information aggregation on the extracted features.

[0010] S3: The pyramid squeezed attention module effectively extracts multi-scale spatial information at a finer-grained level and develops long-term channel dependencies.

[0011] S4: After integrating the information, the features are input into a U-shaped network composed of asymmetric convolution blocks. Asymmetric convolution blocks are used to enhance the effect of local key points. At the same time, the encoding and decoding structure of the U-shaped network can better obtain features of different scales and depths, and enhance the semantic information of remote sensing images.

[0012] S5: The output features are then input into the local and non-local feature extraction module and the pyramid squeeze attention module for final feature enhancement and image reconstruction;

[0013] S6: Optimize the model using the root mean square error loss and adjust the training parameters of the network to obtain the best results.

[0014] As a further optimized solution, further, step S1 is specifically as follows:

[0015] Step S101: performing a bicubic interpolation operation on the multispectral image to upsample it to the same size as the full-color image;

[0016] Step S102: performing preliminary feature extraction on the multispectral image and the panchromatic image separately, and using skip connections to prevent information loss;

[0017] Furthermore, the step S2 is specifically as follows:

[0018] Step S201: concatenate the two features obtained in step 1 and input them into the multi-scale residual block and the large core attention module;

[0019] Step S202: In the multi-scale residual block, convolutions of different sizes are introduced to obtain local features of different scales, and a cascade branch consisting of three 3×3 convolutions is introduced separately to obtain more refined features of the corresponding 5×5 and 7×7 scales. Finally, a 1×1 convolution is used to realize information aggregation of multi-scale features and complement local scale information.

[0020] Step S203: In the large core attention module, in combination with the attention mechanism, spatial long-range convolution space (depth convolution dilated convolution), local convolution (depth convolution) and channel convolution (point convolution) are used to obtain the attention map of local information in space and channel and long-range dependency;

[0021] Step S204: The features extracted by the two branches are preliminarily aggregated using 1×1 convolution, and the features are mapped to a high-dimensional space for subsequent deep feature fusion;

[0022] Furthermore, the step S3 specifically includes:

[0023] Step S301: Integrate the information of the input feature map through a multi-scale pyramid convolution structure. At the same time, by compressing the channel dimension of the input feature tensor, extract spatial information of different scales from each channel-level feature map;

[0024] Step S302: extracting the channel-level attention weight of the multi-scale feature map using the channel attention module, and multiplying it with the overall multi-scale feature map to obtain the final feature map;

[0025] Furthermore, the step S4 specifically includes:

[0026] Step S401: input the features into the asymmetric convolution block, and use asymmetric convolution to enhance the effect of local key points;

[0027] Step S402: using the encoding and decoding structure of the U-type network to better obtain features of different scales and depths, while enhancing the semantic information of the remote sensing image;

[0028] Furthermore, the step S5 specifically includes:

[0029] Step S501: The features output from step S4 are input again into the local and non-local feature extraction module and the pyramid squeeze attention module for final feature enhancement and image reconstruction; in order to maintain the spectral consistency of the generated image, we add the upsampled multispectral image to the final result at the end of the model to obtain the final multispectral image with high spatial resolution;

[0030] Furthermore, the step S6 specifically includes:

[0031] Step S601: Calculate the loss function in the entire fusion process through formula (1) to constrain and optimize network parameters:

[0032]

[0033] Among them, N represents the number of randomly selected training samples, GT i represents the corresponding i-th ground truth image, HRMS i represents the corresponding i-th reconstructed high-resolution multispectral image;

[0034] Step S602: Fine-tune the training parameters of the model, such as learning rate, training batch, etc.

[0035] Furthermore, the simple convolution block is a 9×9 convolution block and two 5×5 convolution blocks.

[0036] Beneficial effects of the present invention:

[0037] 1. Aiming at the fact that the transformer-based fusion method ignores the adaptability of the channel dimension and brings expensive computational costs, the present invention extracts features of remote sensing images from the channel and spatial perspectives by combining local and non-local methods, and constructs a pyramid squeeze attention module to fully integrate the extracted features. It uses a U-shaped feature extraction network based on asymmetric convolution to enhance the semantic information of the features and fully integrate the context, thereby greatly improving the image quality of the multispectral and panchromatic image fusion results.

[0038] 2. Aiming at the differences between local and non-local channel and spatial information, the present invention adopts a pyramid squeezing attention module to fully integrate this information, so that different information can interact and merge well.

[0039] 3. Under the premise of ensuring the stable improvement of the quality of the fusion results, the present invention can perform fusion with a smaller amount of parameters and test time, thereby improving the fusion speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The present invention is a flowchart of a remote sensing image fusion method based on multi-scale feature enhancement of a large kernel attention mechanism.

[0041] Figure 2 This is a principle framework diagram of the remote sensing image fusion method based on multi-scale feature enhancement of the large kernel attention mechanism of the present invention;

[0042] Figure 3 A schematic diagram of a multi-scale residual block of a remote sensing image fusion method for multi-scale feature enhancement based on a large kernel attention mechanism of the present invention;

[0043] Figure 4 A schematic diagram of a large-core attention module of a remote sensing image fusion method for multi-scale feature enhancement based on a large-core attention mechanism of the present invention;

[0044] Figure 5 It is a schematic diagram of a pyramid squeezing attention module of a remote sensing image fusion method for multi-scale feature enhancement based on a large kernel attention mechanism of the present invention;

[0045] Figure 6 A schematic diagram of a U-shaped network based on asymmetric convolution of a remote sensing image fusion method with multi-scale feature enhancement based on a large kernel attention mechanism of the present invention;

[0046] Figure 7 This is a visualization of the fusion results of the remote sensing image fusion method based on the multi-scale feature enhancement of the large kernel attention mechanism on the GaoFen2 dataset. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Example 1

[0049] In order to solve the problem that the transformer-based remote sensing image fusion method cannot capture channel adaptability and the cost of secondary computational complexity is too high for high-resolution remote sensing images, the present invention proposes a remote sensing image fusion method based on multi-scale feature enhancement of a large kernel attention mechanism, the method comprising the following steps:

[0050] Step 1: Preprocess the low-resolution multispectral image and panchromatic image and perform preliminary feature extraction through two simple convolution blocks;

[0051] Step 2: construct a local and non-local feature extraction module consisting of a multi-scale residual block and a large core attention module. The large core attention module is mainly used to obtain long-range dependencies, local spatial information, and adaptability of spatial and spectral dimensions. The multi-scale residual block is mainly used to extract multi-scale local spatial information and perform preliminary information aggregation on the extracted features.

[0052] Step 3: The pyramid squeezed attention module effectively extracts multi-scale spatial information at a finer-grained level and develops long-term channel dependencies.

[0053] Step 4: After integrating the information, the features are input into the U-shaped feature extraction network composed of asymmetric convolution blocks. Asymmetric convolution blocks are used to enhance the effect of local key points. At the same time, the encoding and decoding structure of the U-shaped network can better obtain features of different scales and depths, and enhance the semantic information of remote sensing images.

[0054] Step 5: The output features are then input into the local and non-local feature extraction module and the pyramid squeeze attention module for final feature enhancement and image reconstruction;

[0055] Step 6: Optimize the model using the root mean square error loss and adjust the training parameters of the network to obtain the best results.

[0056] The present invention mainly consists of three parts: a local and non-local feature extraction module composed of a multi-scale residual block and a large core attention module, a pyramid squeeze attention module and a U-type network based on asymmetric convolution. The local and non-local feature extraction module is used to initially extract local and non-local channel and spatial information in multispectral images and panchromatic images. After the extracted features are fused and interacted across channels by the pyramid squeeze attention module, the U-type network based on asymmetric convolution is used to perform deep feature extraction. Finally, the local and non-local feature extraction module and the pyramid squeeze attention module are used again to perform the final image reconstruction to achieve the fusion of remote sensing images.

[0057] Example 2

[0058] like Figure 2 As shown, in order to obtain the initial features of different images, step 1 is optimized, specifically:

[0059] Step 1.1: Input the multispectral image and the panchromatic image into the network, and perform an interpolation upsampling operation on the multispectral image so that the size of the multispectral image is consistent with that of the panchromatic image;

[0060] Step 1.2: Perform simple shallow feature extraction on the upsampled multispectral image and panchromatic image respectively, extract information through 9×9 and two 5×5 convolutions, and use a skip connection to prevent information loss.

[0061] Example 3

[0062] Based on the above embodiment, the features of different images are initially extracted by the convolution block to optimize step 2, specifically:

[0063] Step 2.1: Concatenate the acquired features along the channel dimension and input them into the multi-scale residual block and the large core attention module;

[0064] Step 2.2: In Figure 3 In the multi-scale residual block shown in the figure, convolutions of different sizes are introduced to obtain local features of different scales, and a cascade branch consisting of three 3×3 convolutions is introduced separately to obtain more refined features of the corresponding 5×5 and 7×7 scales. Finally, a 1×1 convolution is used to realize information aggregation of multi-scale features and complement local scale information.

[0065] Step 2.3: In Figure 4 In the large core attention module shown in the figure, we decompose the large core convolution into deep convolution, dilated convolution, deep convolution and point convolution, so as to obtain local information in space and channels and features of long-range dependencies at a low computational cost. At the same time, we combine the attention mechanism to adaptively enhance the acquired features;

[0066] Step 2.4: The features extracted by the two branches are initially aggregated using 1×1 convolution, and the features are mapped to a high-dimensional space for subsequent deep feature fusion;

[0067] Example 4

[0068] Existing transformer-based frameworks mostly use simple convolution fusion operations to fuse local features with long-range features, without considering the differences between different features. Figure 5 The pyramid squeeze attention module shown is used to integrate the various features obtained, specifically:

[0069] Step 3.1: By using grouped convolutions of different sizes in the pyramid structure, features of different spatial resolutions and depths can be generated. By compressing the number of channels of the input features in each branch and establishing cross-channel interactions in a local manner, multi-scale spatial information on each channel-level feature map can be effectively extracted and the representation capability of the features can be enhanced;

[0070] Step 3.2: Then, the output multi-scale feature map is input into the channel attention weight acquisition module along the channel dimension and normalized using the softmax function. Finally, the weight of each channel is concatenated along the channel dimension and multiplied with the overall multi-scale feature map to obtain the final feature map.

[0071] Example 5

[0072] Considering the scale effect of remote sensing images, images of different scales contain different ground object information. We designed Figure 6 The U-shaped network based on asymmetric convolution is shown, specifically:

[0073] Step 4.1: Embed the asymmetric convolution block into the U-shaped network, and use the asymmetric convolution block to enhance the central skeleton of the square convolution kernel, improve the robustness to image rotation distortion, and enhance the influence of local key points;

[0074] Step 4.2: The up-sampling and down-sampling operations in the U-shaped network can deepen the network in an encoding and decoding manner and enrich the semantic information of the features; at the same time, the skip connection of each layer can effectively alleviate the information loss caused by the up-sampling and down-sampling operations;

[0075] Example 6

[0076] In order to complete the final image reconstruction, we extract and fuse the features obtained from the U-shaped network. Specifically:

[0077] Step 5.1: The features output by the U-shaped network are input again into the local and non-local feature extraction module and the pyramid squeeze attention module for final feature enhancement and image reconstruction; in order to maintain the spectral consistency of the generated image, we add the upsampled multispectral image to the final result at the end of the model to obtain the final multispectral image with high spatial resolution;

[0078] Example 7

[0079] Step 6.1: Calculate the loss function of the entire fusion process through formula (1) to constrain and optimize network parameters:

[0080]

[0081] Among them, N represents the number of randomly selected training samples, GT i represents the corresponding i-th ground truth image, HRMS i represents the corresponding i-th reconstructed high-resolution multispectral image;

[0082] Step 6.2: Fine-tune the training parameters of the model, set the network learning rate to 0.01, and the training batch size to 32.

[0083] To prove the effect of the present invention, the following experiment was carried out

[0084] 1. Experimental environment

[0085] The experimental results are as follows Figure 7 As shown, the method of the present invention is further illustrated by experimental imaging:

[0086] The data processing part of this experiment is carried out on the MATLAB R2018b platform, and the PyTorch framework is used for training the learning network model. The computing environment is NVIDIA GeForce RTX 3060GPU and a PC with 16G memory. The algorithms compared in the experiment include two traditional methods, PRACS and MTF-GLP, a variational model method such as the P-XS method, and five deep learning methods, namely the PNN method, PanNet method, MSDCNN method, FusionNet method, and Pansformer method.

[0087] The network was optimized using Adam with a batch size of 1200. Each training batch had 32 images. All deep learning methods were performed on an Nvidia GeForce RTX 3060 GPU with 12-GB RAM, and other methods were implemented in MATLAB 2018b.

[0088] 2. Experimental Dataset

[0089] Furthermore, the GaoFen-2 dataset is selected for experiments. The specific experiments are as follows:

[0090] The spatial resolution of multispectral and panchromatic images in the GaoFen-2 dataset is 4m and 1m. We cropped and downsampled the GaoFen-2 source images according to Wald's protocol, and obtained 8515 sets of multispectral / panchromatic / ground truth image pairs with corresponding sizes of 16 / 64 / 64. All image pairs were divided into 5 / 1 for training / validation. In addition, we separately produced 50 sets of 128 / 512 / 512 multispectral / panchromatic / ground truth image pairs for simulation testing.

[0091] This simulation experiment uses 128 / 512 / 512 multispectral / panchromatic / ground truth images for testing. Figure 7 The visualization pictures of the fusion results of the method of the present invention and the comparative method are given. Figure 7 (a) is the result of fusion of PRACS method; Figure 7 (b) is the result of the fusion of the MTF-GLP method; Figure 7 (c) is the result of the fusion of the P-XS method; Figure 7 (d) is the result of PNN method fusion; Figure 7 (e) is the result of the fusion of the MSDCNN method; Figure 7 (f) is the result of PanNet fusion; Figure 7 (g) is the result of FusionNet method fusion; Figure 7 (h) is the result of the fusion of the Pansformer method; Figure 7 (i) is the result of the fusion of the method of the present invention; Figure 7 (j) is the ground truth image. It can be found that Figure 7 (j) Compared with Figure 7 (a) with Figure 7 (b) There is obvious spatial blur and the image details are completely unclear. Figure 7 (c) Although the overall outline is relatively clear compared to the first two images, the spectrum is severely distorted. Figure 7 (d) with Figure 7 (e) Compared to Figure 7 (c) The overall color and spatial details are slightly improved, but the overall color is obviously blurred. Figure 7 (f) also shows obvious spectral and spatial distortion. Figure 7 (g) with Figure 7 (h) The spatial details of the overall image are relatively close Figure 7 (j), but the colors in its local areas are blurred and the spatial information is slightly over-smoothed. Figure 7(i) The closest in terms of both spatial details and color preservation Figure 7 (j).

[0092] Table 1 shows the performance indicators of the method of the present invention and the comparative method. This experiment adopted the following performance indicators: The correlation coefficient (CC) calculates the similarity of spatial pixels between the fused multispectral image and the original multispectral image. The larger the value of the correlation coefficient, the better the fusion result. ERGAS (ERGAS) represents the difference between the global reflectances of multispectral images. The smaller its value, the better the fusion result. Spectral Angle Mapper (SAM) reflects the difference between the spectral curves of multispectral images. The smaller its value, the better the fusion result. The universal image quality index (UIQI) and its 4-band extended evaluation index Q4 respectively represent the image quality of the original image and the generated result as a whole and the average of the 4 bands. The larger its value, the better the fusion result. The quality evaluation results are shown in Table 1:

[0093] In Table 1, the bold numbers represent the best values ​​of each indicator. From the perspective of fusion objective evaluation indicators, each indicator of the method proposed in the present invention is superior to other methods.

[0094] Table 1 Quantitative evaluation results of different methods on GaoFen-2 dataset

[0095] SAM↓ ERGAS↓ CC↑ UIQI↑ Q4↑ PRACS 3.2445 3.7216 0.8437 0.8277 0.8018 MTF-HPF 2.7758 4.0425 0.8829 0.8630 0.8598 P-XS 4.0312 4.6858 0.8164 0.7123 0.7000 PNN 2.1781 2.3802 0.9719 0.9307 0.9335 MSDCNN 1.8563 2.1614 0.9614 0.9427 0.9424 PanNet 1.5993 2.0313 0.9775 0.9494 0.9440 FusionNet 1.2487 1.6437 0.9766 0.9510 0.9478 Panformer 1.2646 1.3457 0.9789 0.9564 0.9593 The present invention 0.9761 1.3431 0.9811 0.9766 0.9706

[0096] The above experimental results show that the method of the present invention can well maintain spectral information and enhance spatial information by using large-core attention modules and multi-scale residual blocks. The overall spectral and spatial errors are lower than the fusion results of other methods. The method of the present invention strengthens the contextual information of the feature map by cross-scale channel aggregation and U-type network encoding and decoding, thereby improving the overall image quality.

[0097] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, any equivalent changes or modifications made according to the structures, features and principles described in the patent scope of the present invention should be included in the scope of the patent application of the present invention.

Claims

1. A remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism, characterized in that: The method comprises the following steps: S1: Preprocess the low-resolution multispectral image and panchromatic image and perform preliminary feature extraction through two simple convolution blocks; S2: construct a local and non-local feature extraction module consisting of a multi-scale residual block and a large core attention module, wherein the large core attention module is used to obtain long-range dependencies, local spatial information, and adaptability of spatial and spectral dimensions, and the multi-scale residual block is used to extract multi-scale local spatial information and perform preliminary information aggregation on the extracted features; The step S2 is specifically as follows: S201: concatenate the two features obtained in step 1 and input them into the multi-scale residual block and the large core attention module; S202: In the multi-scale residual block, convolutions of different sizes are introduced to obtain local features of different scales, and a cascade branch consisting of three 3×3 convolutions is introduced separately to obtain more refined features of the corresponding 5×5 and 7×7 scales. Finally, a 1×1 convolution is used to realize information aggregation of multi-scale features and complement local scale information. S203: In the large core attention module, the attention mechanism is combined to use spatial long-range convolution space, local convolution and channel convolution to obtain the attention map of local information in space and channel and long-range dependency; S204: The features extracted by the two branches are preliminarily aggregated using 1×1 convolution, and the features are mapped to a high-dimensional space for subsequent deep feature fusion; S3: Effectively extract multi-scale spatial information at a finer-grained level and develop long-term channel dependencies through the pyramid squeezed attention module; The step S3 specifically includes: S301: Integrate the information of the input feature map through a multi-scale pyramid convolution structure. At the same time, extract spatial information of different scales from each channel-level feature map by compressing the channel dimension of the input feature tensor. S302: extracting the channel-level attention weight of the multi-scale feature map by using the channel attention module, and multiplying it with the overall multi-scale feature map to obtain the final feature map; S4: After integrating the information, the features are input into the asymmetric convolutional blocks. U In the network, asymmetric convolution blocks are used to enhance the effect of local key points. U The encoding and decoding structure of the network can obtain features of different scales and depths, while enhancing the semantic information of remote sensing images; S5: The output features are then input into the local and non-local feature extraction module and the pyramid squeeze attention module for final feature enhancement and image reconstruction; S6: Optimize the model using the root mean square error loss and adjust the training parameters of the network to obtain the best results.

2. The remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism according to claim 1 is characterized in that: The step S1 is specifically as follows: S101: performing a bicubic interpolation operation on the multispectral image to upsample it to the same size as the panchromatic image; S102: Perform separate preliminary feature extraction on the multispectral image and the panchromatic image respectively, and use skip connections to prevent information loss.

3. The remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism according to claim 1 is characterized in that: The step S4 specifically includes: S401: Input the features into the asymmetric convolution block and use asymmetric convolution to enhance the effect of local key points; S402: Utilize the encoding and decoding structure of the U-type network to obtain features of different scales and depths, while enhancing the semantic information of the remote sensing image.

4. The remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism according to claim 1 is characterized in that: The S5 specifically includes: S501: The features output from step S4 are input again into the local and non-local feature extraction module and the pyramid squeezing attention module for final feature enhancement and image reconstruction; in order to maintain the spectral consistency of the generated image, the upsampled multispectral image is added to the final result at the end of the model to obtain the final multispectral image with high spatial resolution.

5. The remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism according to claim 1 is characterized in that: The step S6 specifically includes: S601: Calculate the loss function of the entire fusion process through formula (1) to constrain and optimize network parameters: （1）； Among them, among them, represents the number of randomly selected training samples, Indicates the corresponding ground truth images, Indicates the corresponding A reconstructed high-resolution multispectral image; S602: Fine-tune the training parameters of the model, where the training parameters include a learning rate and a training batch.

6. The remote sensing image fusion method based on multi-scale feature enhancement of large kernel attention mechanism according to claim 2 is characterized in that: The simple convolution block is a 9×9 and two 5×5 convolutions.

Images