Lightweight remote sensing image super-resolution reconstruction method

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a lightweight remote sensing image super-resolution reconstruction method, which comprises the following steps of: 1, preparing a data set: acquiring a remote sensing image data set, and dividing the remote sensing image data set into a training set, a verification set and a test set according to a proportion; 2, constructing a network model: constructing three parts including a shallow feature extraction module, a deep feature extraction module and a reconstruction module, wherein the deep feature extraction part comprises a lightweight multi-feature fusion block and a fusion enhancement module; and step 3, training the network model: training the super-resolution reconstruction network by using the loss function until the training frequency reaches an initial set threshold value or the value of the loss function reaches a preset range, and completing the training of the network model. According to the method, the structure and the training mode of the lightweight super-resolution reconstruction network are designed, and lightweight local and global features are introduced into the channel dimension and the space dimension respectively, so that the overall super-resolution reconstruction network maintains the quality of the super-resolution image while reducing the model scale.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a lightweight remote sensing image super-resolution reconstruction method. Background Art

[0002] Super-resolution (SR) technology for remote sensing images is a technique that converts low-resolution (LR) images into high-resolution (HR) images. Its purpose is to increase image resolution, resulting in clearer, more detailed, and more realistic images. Compared to natural images, remote sensing images have a wider field of view, can capture a wider range of ground information, and contain richer texture information. However, due to the large volume of remote sensing image data, processing requires significant computing resources and time. Furthermore, in some cases, due to limitations in sensor performance, data transmission, and storage, the resolution of remote sensing images may be low, making it difficult to present sufficient detail and clarity, thus limiting their practical applications. Super-resolution technology for remote sensing images can generate a high-resolution image from multiple low-resolution images, significantly enhancing the detail and clarity of remote sensing images. This technology is crucial for improving the analysis accuracy, object recognition capabilities, and monitoring effectiveness of remote sensing images, and can be widely used in areas such as land cover classification, building detection, and three-dimensional reconstruction.

[0003] Chinese patent publication number CN113034361A, titled "A Method for Super-Resolution Reconstruction of Remote Sensing Images Based on an Improved ESRGAN," proposes improvements to the remote sensing image super-resolution reconstruction network, comprising a generator and a discriminator. The generator network consists of 64 3×3 convolutional layers, 23 RRDB modules, and a LeakyReLU activation function. The discriminator network, comprised of six layers, employs a fully convolutional network with even-sized kernels, interleaved with batch normalization (BN) layers and LeakyReLU activation layers. The first layer of the discriminator network receives as input the original low-resolution remote sensing image realA, the image upscaled by bicubic interpolation, and the channel-merged fakeB output from the generator network. By alternately training the generator and discriminator networks and updating their parameters, the improved remote sensing image super-resolution reconstruction network model is achieved. This approach achieves better super-resolution reconstruction quality by increasing network depth, but suffers from redundant feature extraction, low efficiency, and excessive parameterization, making the super-resolution reconstruction network difficult to implement in practical deployments.

[0004] In summary, how to solve the problem that the current remote sensing image super-resolution reconstruction method has a large network model and is difficult to deploy in practice is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In view of the shortcomings of the existing technology, the present invention provides a lightweight remote sensing image super-resolution reconstruction method, which solves the problems raised in the above background technology.

[0007] (2) Technical solution

[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0009] A lightweight remote sensing image super-resolution reconstruction method includes the following steps:

[0010] Step 1, prepare the data set: obtain the remote sensing image data set and divide it into training set, validation set, and test set in proportion;

[0011] Step 2: Build a network model: The model consists of three parts: shallow feature extraction, deep feature extraction, and reconstruction modules. The deep feature extraction part includes a lightweight multi-feature fusion block and a fusion enhancement module.

[0012] Step 3: Train the network model: Use the loss function to train the super-resolution reconstruction network until the number of training times reaches the initial set threshold or the value of the loss function reaches the preset range, then the network model training is completed;

[0013] Step 4: Fine-tune the network model: Use the validation set to adjust the network model and optimize the network model parameters;

[0014] Step 5: Solidify the network model: After fine-tuning is completed, solidify the fine-tuned network parameters to determine the final remote sensing image super-resolution reconstruction network parameters.

[0015] Furthermore, the network model in step 2 includes a convolutional layer for shallow feature extraction, four lightweight multiple feature fusion blocks, a fusion enhancement module and a reconstruction module. A global residual connection is introduced before the reconstruction module, and a jump connection after bicubic upsampling is introduced after the reconstruction module.

[0016] Furthermore, the convolution layer for shallow feature extraction adopts a 3×3 convolution, the lightweight multiple feature fusion block includes three channel space cross-cascade blocks, a fusion enhancement module and a context extraction module, and the reconstruction module adopts a convolution layer with a convolution kernel of 3×3 and a PixelShuffle pixel reorganization operation. The convolution layer adjusts the number of channels of the input features, and the PixelShuffle pixel reorganization operation reconstructs the feature map.

[0017] Furthermore, the channel-space cross-cascade block is composed of a spatial attention interaction block, a channel attention interaction block, a grouped residual shuffling block, a channel self-attention mechanism and a spatial self-attention mechanism. The spatial attention interaction block includes a 1×1 convolution layer, an R-type function and a Sigmoid function for extracting spatial information; the channel attention interaction block contains grouped convolution and a 1×1 convolution layer, an R-type function and a Sigmoid function for extracting channel information; the grouped residual shuffling block includes grouped convolution, an R-type function and a channel shuffling operation to break the information isolation between groups and enhance feature interaction. The channel self-attention mechanism and the spatial self-attention mechanism are used to extract global features in the channel dimension and the spatial dimension.

[0018] Furthermore, the context extraction module includes a 1×1 convolution layer, strided convolution, maximum pooling operation, grouped convolution, bicubic upsampling operation and Sigmoid function. By introducing the context extraction module, the spatial details in the fusion function are expanded, effectively enhancing the model performance.

[0019] Furthermore, the fusion enhancement module includes a 1×1 convolution layer, an R-type function, a self-calibration group convolution module and a contrast channel attention module. The module is used below to represent a lightweight multiple feature fusion block or a channel space cross cascade block. The outputs of the first module and the second module are respectively passed through a 1×1 convolution layer and an R-type function to perform channel splicing on the output as a new module output and the next module is respectively passed through a 1×1 convolution layer and an R-type function to perform channel splicing, and so on. Then, the output is passed through the self-calibration group convolution module and the contrast channel attention module in parallel to obtain the output. The channel concatenation finally passes through a 1×1 convolution layer, and the input features are fused and enhanced through contrastive channel attention and self-calibration group convolution. The self-calibration group convolution block includes a maximum pooling operation, grouped convolution, channel shuffling, upsampling, a Sigmoid function, and a grouped residual shuffle block to extract features at different scales and realize the interaction of complex features. The contrastive channel attention includes a contrast operation, a 1×1 convolution, an R-type function, and a Sigmoid function. The channel weight is generated by calculating the mean and standard deviation of each channel to more comprehensively describe the channel.

[0020] Furthermore, step 3 also includes selecting a loss function and determining an evaluation index, using pixel loss as the loss function, and using value signal-to-noise ratio, structural similarity and number of parameters as evaluation indexes.

[0021] (3) Beneficial effects

[0022] Compared with the existing technology, the present invention provides a lightweight remote sensing image super-resolution reconstruction method with the following beneficial effects:

[0023] The present invention improves the performance of the model while reducing the parameters of the model by designing a lightweight remote sensing super-resolution reconstruction network. The lightweight super-resolution reconstruction network is used in the field of remote sensing super-resolution to solve the problem that the existing network model is large and difficult to deploy in practice. By designing a channel-space cross-cascade block, the information interaction between channel attention and spatial attention is enhanced from both local and global aspects, and the features are effectively extracted and enhanced while maintaining the lightweight nature of the model. By designing a fusion enhancement module, multi-level features are integrated in series through the sequence of near and far distances, which not only obtains more complete features but also enhances the ability to represent complex features. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method steps of the present invention;

[0025] Figure 2 Schematic diagram of the overall super-resolution reconstruction network structure in the method of the present invention;

[0026] Figure 3 Schematic diagram of the structure of the lightweight multi-feature fusion block in the method of the present invention;

[0027] Figure 4 Schematic diagram of the structure of the channel space cross cascade block in the method of the present invention;

[0028] Figure 5 Schematic diagram of the structure of the spatial attention interaction block in the method of the present invention;

[0029] Figure 6 Schematic diagram of the structure of the channel attention interaction block in the method of the present invention;

[0030] Figure 7 Schematic diagram of the structure of the grouped residual shuffling block in the method of the present invention;

[0031] Figure 8 Schematic diagram of the structure of the context extraction module in the method of the present invention;

[0032] Figure 9 Schematic diagram of the structure of the fusion enhancement module in the method of the present invention;

[0033] Figure 10 Schematic diagram of the structure of the self-calibration group convolution block in the method of the present invention;

[0034] Figure 11 Schematic diagram of the structure of the comparison channel attention module in the method of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] Example

[0037] like Figure 1-11 As shown, a lightweight remote sensing image super-resolution reconstruction method proposed in one embodiment of the present invention includes the following steps:

[0038] Step 1, prepare the dataset;

[0039] The first remote sensing dataset, the AID dataset, contains 10,000 images covering 30 scene categories (airport, bare land, baseball field, beach, bridge, center, church, commercial, densely populated, desert, farmland, forest, industrial, grassland, medium-sized residential, mountain, park, parking lot, playground, pond, harbor, train station, resort, river, school, sparsely populated, square, stadium, storage tank, and viaduct). Each category has approximately 200-420 images, and the image size is 600 pixels × 600 pixels. The dataset was cleaned to contain 5,000 images, which were then split into training and validation sets in a 7:3 ratio. The second remote sensing dataset, the UCMLU dataset, was prepared. The image size is 256 pixels × 256 pixels, and it contains 21 scene categories, with 100 images per category, for a total of 2,100 images. The dataset was cleaned to contain 1,500 images, which were used as the test set. The cleaned dataset is bicubic downsampled 4 times to obtain low-resolution images and construct the paired datasets required for training, validation, and testing.

[0040] Perform image augmentation on the images in the dataset in step 1. For the same pair of images, the high-resolution image is randomly flipped and cropped to a size of 256×256. The same operation is performed on the low-resolution image to obtain an input image size of 64×64, which serves as the input of the entire network. The random size and position can be achieved through software algorithms. Processing the images in the dataset through image augmentation is to enhance the robustness of the network and improve the network generalization ability.

[0041] Step 2: Build a network model; build a lightweight remote sensing image super-resolution reconstruction network.

[0042] Super-resolution reconstruction network models, such as Figure 2As shown in Figure 1, the super-resolution reconstruction network consists of convolutional layer 1, lightweight multi-feature fusion block 1, lightweight multi-feature fusion block 2, lightweight multi-feature fusion block 3, lightweight multi-feature fusion block 4, fusion enhancement module, reconstruction module and bicubic upsampling. The convolution kernel size of convolutional layer 1 is 3×3, the stride is 1, and the padding is 1, which is used to extract shallow features. The low-resolution image is enlarged to the target size through bicubic upsampling and is residually connected with the output of the reconstruction module to compensate for the important details of the low-level features. This process can be expressed as follows:

[0043] I SR =H re (I LR )+H bicubic (I LR )

[0044] Among them, I LR Represents the input low-resolution image, I SR Represents the output super-resolution image, H re (·) represents the main function of the super-resolution reconstruction network, H bicubic (·) denotes the bicubic upsampling operation.

[0045] Lightweight multi-feature fusion blocks, such as Figure 3 As shown in the figure, it is used for deep feature extraction and consists of three channel space cross cascade blocks, a fusion enhancement module and a context extraction module. The specific structure of the channel space cross cascade block is as follows: Figure 4As shown, first use the channel segmentation technology to divide the features into X1 and X2. X1 passes through the spatial attention interaction block and X2 passes through the grouped residual shuffle block for channel splicing to obtain X3. X2 passes through the grouped residual shuffle block and the channel attention interaction block and X1 for channel splicing to obtain X4. X3 passes through the 1×1 convolution layer and the grouped residual shuffle block to obtain X5. X4 passes through the 1×1 convolution layer to obtain X6. X5 passes through the spatial attention interaction block and the output is spliced with X6 to obtain X8. X6 passes through the channel attention interaction block and the output is spliced with X5 to obtain X7. X7 and X8 pass through the 1×1 convolution layer respectively and then pass through the 1×1 convolution layer to obtain the local feature FL. The local feature FL is first divided into X9 and X10 through channel segmentation. X9 is transformed from C×H×W to (HW)×C through dimension transformation, and then linearly transformed to obtain Q (query), K (key) and V (value) ) feature matrix, the Q feature matrix and the K feature matrix are multiplied, and the spatial dimension attention weight matrix is calculated by the softmax function. The spatial dimension attention weight matrix is multiplied by the V feature matrix to obtain the V1 feature matrix, and then the Q, K, and V1 feature matrices are converted by dimension (the dimension is converted from (HW)×C to C×(HW)) to obtain the Q′, K′, and V′ feature matrices. The Q′ feature matrix and the K′ feature matrix are multiplied, and the channel dimension attention weight matrix is calculated using softmax. The channel dimension attention weight matrix is multiplied by the V′ feature matrix to obtain the attention output sequence. Then, the output is converted by dimension (the dimension is converted from C×(HW) to C×H×W) and the channel is spliced with X10 to obtain the final output (that is, the output of the channel cross-cascade block); the convolution kernel size of all convolutional layers is 1×1, the stride is 1, and the padding is 0. The spatial self-attention and channel self-attention can be designed as:

[0046]

[0047] in, is the scaling factor, K T is the transposed matrix of the K matrix, K′ T is the transposed matrix of the K′ matrix.

[0048] Spatial attention interaction blocks, such as Figure 5 As shown in , it includes convolution layer 1, R-type function, convolution layer 2 and Sigmoid function. The convolution kernel size of convolution layer 1 and convolution layer 2 is 1×1, the step size is 1, and the padding is 0, which is used to extract spatial information; channel attention interaction block, such as Figure 6As shown, it includes group convolution, convolution layer 1, R-type function, convolution layer 2 and Sigmoid function. The convolution kernel size of group convolution is 3×3, with a step size of 1 and a padding of 1. The convolution kernel size of convolution layer 1 and convolution layer 2 is 1×1, with a step size of 1 and a padding of 0, which is used to extract channel information; group residual shuffling block, as shown Figure 7 As shown in , it includes group convolution, R-type function and channel shuffling operation. The convolution kernel size of group convolution is 3×3, the step size is 1, and the padding is 1, which breaks the information isolation between groups and enhances feature interaction; the context extraction module, such as Figure 8 As shown in the figure, it includes convolution layer 1, convolution layer 2, convolution layer 3, strided convolution, maximum pooling operation, grouped convolution, bicubic upsampling operation and Sigmoid function. The convolution kernel size of convolution layer 1, convolution layer 2 and convolution layer 3 is 1×1, the stride is 1, and the padding is 0. The convolution kernel size of strided convolution is 3×3, the stride is 2, and the padding is 1. The convolution kernel size of grouped convolution is 3×3, the stride is 1, and the padding is 1. By introducing the context extraction module, the spatial details in the fusion function are expanded, and the model performance is effectively enhanced.

[0049] Fusion enhancement modules, such as Figure 9 As shown in the figure, it includes a 1×1 convolution layer, an R-type function, a self-calibration group convolution module, and a contrast channel attention module. The module is used below to represent a lightweight multiple feature fusion block or a channel space cross-cascade block. The outputs of the first and second modules are respectively passed through a 1×1 convolution layer and an R-type function to perform channel splicing on the output as the new module output. The next module is passed through a 1×1 convolution layer and an R-type function to perform channel splicing again, and so on. The process can be expressed as follows:

[0050] ReLU(Conv 1×1 (F i )), i=1,

[0051] Conv 1×1 (Concat(T i-1 ,ReLU(Conv 1×1 (F i ))), i=2,……,n-1,

[0052] Concat(T i-1 ,ReLU(Conv 1×1 (F i ))), i=n,

[0053] F out =T i ,

[0054] Among them, F irepresents the output feature of the i-th module, ReLU(·) represents the R-type function, Concat(·) represents the channel concatenation operation, F out =T i Represents the serial output of i modules, Conv 1×1 Represents a 1×1 convolutional layer.

[0055] The output then passes through the self-calibration group convolution module and the contrastive channel attention module in parallel. The obtained output is channel-joined and finally passes through a 1×1 convolution layer. The input features are fused and enhanced through contrastive channel attention and self-calibration group convolution. The convolution kernel size of the 1×1 convolution layer is 1×1, the stride is 1, and the padding is 0.

[0056] Self-calibrated group convolution blocks, such as Figure 10 As shown in the figure, it includes maximum pooling operation, grouped convolution, channel shuffling, upsampling, Sigmoid function and grouped residual shuffling block. The convolution kernel size of the grouped convolution is 3×3, the step size is 1, and the padding is 1. Features are extracted at different scales to achieve the interaction of complex features.

[0057] Compared with the channel attention module, Figure 11 As shown in the figure, it includes contrast operation, 1×1 convolution, R-type function and Sigmoid function. The convolution kernel size of 1×1 convolution is 1×1, the step size is 1, and the padding is 0. The channel weight is generated by calculating the mean and standard deviation of each channel to describe the channel more comprehensively.

[0058] The reconstruction module includes a 3×3 convolutional layer and a PixelShuffle pixel reorganization operation. The convolution kernel size of the convolutional layer is 3×3, the stride is 2, and the padding is 1.

[0059] In order to ensure the robustness of the network, retain more structural information, and fully extract image features, the present invention uses an activation function, which is an R-type function. The definition of the R-type function is as follows:

[0060]

[0061] Step 3: Train the network model. First, select a suitable loss function and determine the evaluation index. Based on the network model in step 2, select a suitable loss function to minimize the difference between the super-resolution image reconstructed by the network and the high-resolution image, and determine the evaluation index to evaluate the performance of the network. The loss function used by the super-resolution reconstruction network is pixel loss.

[0062] In the supervised image super-resolution task, in order to make the generated image (SR) as close as possible to the real high-resolution image (HR), the L1 loss is used to calculate the error between the values of the corresponding pixel positions of SR and HR, that is, the pixel loss. The formula is as follows:

[0063]

[0064] Where x is the input low-resolution image, y is the high-resolution image, and F(·) is the super-resolution reconstruction network.

[0065] The evaluation metrics used are peak signal-to-noise ratio (PSNR), structural similarity, and number of parameters. The PSNR measures the difference between the super-resolved image and the ground-truth image; a higher PSNR indicates better super-resolved image quality. Structural similarity measures contour preservation during the super-resolved process by calculating the structural similarity between the super-resolved image and the ground-truth image; a higher structural similarity indicates a smaller difference. The number of parameters describes the total number of trainable parameters in the model and measures the size of the model. The definitions of PSNR and structural similarity are as follows:

[0066]

[0067] Among them, μ x , μ y represent the mean of image x and y respectively, and Represents the standard deviation of image x and y, σ xy represents the covariance of images x and y, and C1 and C2 are constants.

[0068] The number of parameters is calculated as follows:

[0069] Fully connected layer: number of parameters = number of input units + number of output units.

[0070] Convolution layer: number of parameters = number of convolution kernels × (convolution kernel height × convolution kernel width × number of input channels) + number of convolution kernels.

[0071] Secondly, train the remote sensing super-resolution reconstruction network model: use the loss function to start training the network until the number of training times reaches the set threshold or the value of the loss function reaches the set range, then the model parameters are considered to have been trained and saved; select the test data set of data set 1 to test the network model, and the model performance is evaluated by the evaluation indicators selected above.

[0072] All experiments were conducted on the AutoDL cloud server, using an NVIDIA RTX 3080Ti 12GB GPU for algorithm acceleration. The training cycle was set to 200 epochs, and the learning rate for the super-resolution reconstruction network was set to 1e-4. The Adam optimizer was chosen as the network parameter optimizer, which has the advantages of simple implementation, high computational efficiency, low memory requirements, and parameter updates that are not affected by gradient scaling, resulting in relatively stable parameters. When the loss stabilizes, the network is considered to be basically trained.

[0073] Step 4: Fine-tune the network model; use the validation set of dataset 1 in step 1 to adjust the network, optimize the network model parameters, and further improve the performance of the super-resolution reconstruction network.

[0074] Step 5: Solidify the network model. After the fine-tuning in step 4 is completed, solidify the fine-tuned network parameters and determine the final super-resolution reconstruction network parameters. When subsequent work requires super-resolution reconstruction of remote sensing images, directly use the low-resolution remote sensing image as the input of the super-resolution reconstruction network to obtain better super-resolution results.

[0075] Among them, the implementation of convolution, activation function, feature channel splicing operation, pixel-level addition operation, pixel-level multiplication operation, feature dimension conversion operation, maximum pooling operation, contrast operation, etc. are algorithms well known to those skilled in the art, and the specific processes and methods can be found in the corresponding textbooks or technical literature.

[0076] The present invention constructs a lightweight remote sensing image super-resolution reconstruction method, which can directly super-reconstruct low-resolution remote sensing images into high-resolution images.

[0077] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A lightweight remote sensing image super-resolution reconstruction method, characterized by: The following steps are involved: Step 1, prepare the data set: obtain the remote sensing image data set and divide it into training set, validation set, and test set in proportion; Step 2: Build a network model: The model consists of three parts: shallow feature extraction, deep feature extraction, and reconstruction modules. The deep feature extraction part includes a lightweight multi-feature fusion block and a fusion enhancement module. Step 3: Train the network model: Use the loss function to train the super-resolution reconstruction network until the number of training times reaches the initial set threshold or the value of the loss function reaches the preset range, then the network model training is completed; Step 4: Fine-tune the network model: Use the validation set to adjust the network model and optimize the network model parameters; Step 5: Solidify the network model: After fine-tuning is completed, solidify the fine-tuned network parameters to determine the final remote sensing image super-resolution reconstruction network parameters.

2. The lightweight remote sensing image super-resolution reconstruction method according to claim 1, characterized in that: The network model in step 2 includes a convolutional layer for shallow feature extraction, four lightweight multi-feature fusion blocks, a fusion enhancement module and a reconstruction module. A global residual connection is introduced before the reconstruction module, and a jump connection after bicubic upsampling is introduced after the reconstruction module.

3. The lightweight remote sensing image super-resolution reconstruction method according to claim 2, characterized in that: The convolution layer for shallow feature extraction uses a 3×3 convolution. The lightweight multiple feature fusion block includes three channel space cross-cascade blocks, a fusion enhancement module and a context extraction module. The reconstruction module uses a convolution layer with a convolution kernel of 3×3 and a PixelShuffle pixel reorganization operation. The convolution layer adjusts the number of channels of the input features, and the PixelShuffle pixel reorganization operation reconstructs the feature map.

4. The lightweight remote sensing image super-resolution reconstruction method according to claim 3, characterized in that: The channel-space cross-concatenation block consists of a spatial attention interaction block, a channel attention interaction block, a grouped residual shuffling block, a channel self-attention mechanism, and a spatial self-attention mechanism. The spatial attention interaction block includes a 1×1 convolution layer, an R-type function, and a Sigmoid function for extracting spatial information; the channel attention interaction block contains grouped convolution and a 1×1 convolution layer, an R-type function, and a Sigmoid function for extracting channel information; the grouped residual shuffling block includes grouped convolution, an R-type function, and a channel shuffling operation to break the information isolation between groups and enhance feature interaction. The channel self-attention mechanism and the spatial self-attention mechanism are used to extract global features in the channel dimension and the spatial dimension.

5. The lightweight remote sensing image super-resolution reconstruction method according to claim 3, characterized in that: The context extraction module includes a 1×1 convolution layer, strided convolution, a maximum pooling operation, a grouped convolution, a bicubic upsampling operation, and a sigmoid function. By introducing the context extraction module, the spatial details in the fusion function are expanded, thereby effectively enhancing the model performance.

6. The lightweight remote sensing image super-resolution reconstruction method according to claim 2, characterized in that: The fusion enhancement module includes a 1×1 convolution layer, an R-type function, a self-calibration group convolution module, and a contrast channel attention module. The module is used below to represent a lightweight multiple feature fusion block or a channel space cross-cascade block. The outputs of the first module and the second module are respectively passed through a 1×1 convolution layer and an R-type function to perform channel splicing on the output as the new module output and the next module is respectively passed through a 1×1 convolution layer and an R-type function for channel splicing, and so on. Then, the self-calibration group convolution module and the contrast channel attention module are passed in parallel to perform channel splicing on the output. Finally, a 1×1 convolution layer is passed to perform fusion enhancement on the input features through contrast channel attention and self-calibration group convolution. The self-calibration group convolution block includes a maximum pooling operation, grouped convolution, channel shuffling, upsampling, a Sigmoid function, and a grouped residual shuffling block, which extracts features at different scales and realizes the interaction of complex features. The contrast channel attention includes a contrast operation, a 1×1 convolution, an R-type function, and a Sigmoid function. The weight of the channel is generated by calculating the mean and standard deviation of each channel to more comprehensively describe the channel.

7. The lightweight remote sensing image super-resolution reconstruction method according to claim 1, characterized in that: The step 3 also includes selecting a loss function and determining an evaluation index, using pixel loss as the loss function, and using value signal-to-noise ratio, structural similarity and number of parameters as evaluation indexes.

Images