A method for enhancing remote sensing mapping images

Abstract

The present invention proposes a method for enhancing remote sensing mapping images, which relates to the field of image enhancement. The present invention proposes a process for enhancing remote sensing mapping images, including the production of remote sensing mapping image datasets, the construction of an initialization module, the construction of an optimized sampling module I, the construction of an optimized sampling module II, the construction of a feature processing module, the construction of a remote sensing mapping image enhancement model, the training and use of the remote sensing mapping image enhancement model; at the same time, a remote sensing mapping image enhancement model is proposed. Among them, the optimized sampling module I effectively enhances the detail performance and the accuracy of local features of remote sensing images through convolution operations and the calculation of compensated feature maps. The optimized sampling module II further improves the correlation between adjacent pixels and the detail restoration effect in remote sensing images by calculating the compensated feature maps through weighted averaging. The remote sensing mapping image enhancement model significantly improves the detail restoration and overall quality of low-quality remote sensing mapping images through the combined optimization of multiple modules.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and particularly relates to a method for enhancing remote sensing mapping images. Background Art

[0002] Remote sensing mapping image enhancement is an important technical means to improve image quality and usability, and is widely used in fields such as geographic information systems, environmental monitoring, and resource exploration. Due to the complex acquisition environment of remote sensing images, they are often affected by atmospheric interference, changes in lighting conditions, and sensor noise, resulting in poor image quality. Through image enhancement technology, the clarity, contrast, and detail performance of images can be improved, the recognition ability of ground object targets can be enhanced, and a reliable basis can be provided for subsequent image processing and analysis.

[0003] Image enhancement technology includes two major categories of methods: spatial domain and frequency domain. Spatial domain methods mainly adjust pixel values to improve image brightness and contrast, such as histogram equalization, gamma correction, etc. Frequency domain methods perform spectral analysis on images through means such as Fourier transform, filter out noise and irrelevant information, and enhance the performance of useful information. In the field of remote sensing mapping, image enhancement technology is usually optimized according to specific application requirements and different ground object characteristics and task objectives.

[0004] With the introduction of advanced technologies such as deep learning, image enhancement methods based on deep neural networks have gradually become a research hotspot. These methods can automatically achieve efficient enhancement of images by learning the characteristics of a large amount of remote sensing image data. Compared with traditional methods, deep learning technology has stronger adaptability and robustness, and can achieve better image enhancement effects in complex remote sensing scenarios, providing strong support for the accuracy and efficiency of remote sensing mapping. Summary of the Invention

[0005] The present invention provides a method for enhancing remote sensing mapping images, aiming to propose a remote sensing mapping image enhancement model. The initialization module provides various types of images for subsequent operations. The first optimized sampling module effectively enhances the detail performance and the accuracy of local features of remote sensing images through convolution operations and the calculation of compensated feature maps. The second optimized sampling module further improves the correlation between adjacent pixels and the detail restoration effect in remote sensing images by calculating the weighted average of compensated feature maps. The remote sensing mapping image enhancement model significantly improves the detail restoration and overall quality of low-quality remote sensing mapping images through the combined optimization of multiple modules.

[0006] The present invention aims to propose a remote sensing mapping image enhancement model and provides a method for enhancing remote sensing mapping images, including the following steps:

[0007] S1. Production of a remote sensing mapping image dataset: Collect multiple high-quality remote sensing mapping images and produce a remote sensing mapping image dataset;

[0008] S2. Construct an initialization module, including 3-fold downsampling and 3-fold upsampling;

[0009] S3. Construct an optimized sampling module I, including 3×3 depth convolution and calculating a type of compensated feature map;

[0010] S4. Construct an optimized sampling module II, including 3×3 depth convolution and calculating a type II compensated feature map;

[0011] S5. Construct a feature processing module, using a symmetric coding structure;

[0012] S6. Construct a remote sensing mapping image enhancement model, including input, initialization module, optimized sampling module I, optimized sampling module II, 3×3 convolution, feature processing module, upsampling, output;

[0013] S7. Train and use the remote sensing mapping image enhancement model. Use the remote sensing mapping image dataset to train the remote sensing mapping image enhancement model. After training, input a single low-quality remote sensing mapping image into the remote sensing mapping image enhancement model, and the remote sensing mapping image enhancement model outputs a single high-quality remote sensing mapping image.

[0014] Preferably, in step S1, for the remote sensing mapping image dataset, collect multiple high-quality remote sensing mapping images, perform downsampling, transposed convolution operation, Gaussian blur, and motion blur on the multiple high-quality remote sensing mapping images to obtain multiple low-quality remote sensing mapping images, and form multiple training pairs with the multiple high-quality remote sensing mapping images to form a remote sensing mapping image dataset.

[0015] Preferably, in step S2, for the initialization module, input a single low-quality remote sensing mapping image I, I ∈ R h ×w×3 , where h, w, and 3 represent the height, width, and channels of the single low-quality remote sensing mapping image I, and R represents the real number space, that is, h and w take real values. Perform 3-fold downsampling on the single low-quality remote sensing mapping image I to obtain a single 3-fold downsampled low-quality remote sensing mapping image I d , which is one-third of h and w. Perform 3-fold upsampling on the single 3-fold downsampled low-quality remote sensing mapping image I d to obtain a single lossy low-quality remote sensing mapping image I u , I u ∈ R h×w×3 , and use the single low-quality remote sensing mapping image I to perform element-wise subtraction on the single lossy low-quality remote sensing mapping image I u to obtain a loss image I loss , I loss ∈ R h×w×3 , I loss= I - I u , '-' represents element-wise subtraction, a single low-quality remote sensing mapping image I, a single lossy low-quality remote sensing mapping image I u , the loss image I loss , the low-quality remote sensing mapping image I d is the output of the initialization module.

[0016] Preferably, in step S2, the initialization module adopts a strategy of 3-fold downsampling and 3-fold upsampling. By downsampling the low-quality remote sensing mapping image, the noise and redundant information in the remote sensing mapping image are reduced, and then the size of the remote sensing mapping image is restored through upsampling. While maintaining the key information of the remote sensing mapping image, the processing complexity is reduced, providing various types of inputs for subsequent operations, including the low-quality remote sensing mapping image, the lossy low-quality remote sensing mapping image, the loss image, and the low-quality remote sensing mapping image.

[0017] Preferably, in step S3, for the optimization sampling module 1, a single low-quality remote sensing mapping image I and a single lossy low-quality remote sensing mapping image I obtained from the output of the initialization module u , the loss image I loss , I ∈ R h×w×3 , I u ∈ R h×w×3 , I loss ∈ R h×w×3 , h, w, and 3 represent the height, width, and channels of a single low-quality remote sensing mapping image I, a single lossy low-quality remote sensing mapping image I u , the loss image I loss , and then the feature F OS1 and the feature F OS2 , F OS1 = DWConv 3×3 (I),

[0018] F OS2 = DWConv 3×3 (I u ), DWConv 3×3 (Iu) represents a 3×3 depth convolution, h1, w1, and c1 represent the height, width, and channels of the feature F OS1 and the feature F OS2 , and then through the feature F OS1 and the feature F OS2 at each pixel position (i, j) and each color channel c, calculate the similarity Similar OS1 between the feature F OS2 and the feature F i，j，c ,

[0019] Represent feature F OS1 For each pixel position (i, j) and each color channel c Represent feature F OS2 For each pixel position (i, j) and each color channel c, i ∈ [0, h1 - 1], j ∈ [0, w1 - 1], c ∈ [0, c1 - 1], · represents dot product Represent The L2 norm of Represent The L2 norm of, and then calculate a class of compensation feature maps C1, C1 ∈ R h×w×3 , Where TConv represents transposed convolution Represents an adaptive hyperparameter for adjusting the compensation intensity, target size = h × w × 3 represents the output target dimension of the transposed convolution TConv as h × w × 3, concatenate a class of compensation feature maps C1 and the loss image I loss On the channel dimension to obtain the mixed image I OSC , I OSC ∈ R h ×w×6 , and then obtain the optimized encoded image I CE1 ,

[0020] I CE1 = Down(DWConv 3×3 (Concat(Linear(I OSC ), I))), where Linear represents the linear layer, Concat represents concatenation on the channel dimension, DWConv 3×3 Represents a 3×3 depth convolution, Down represents downsampling, and the optimized encoded image I CE1 Is the output of the first optimized sampling module

[0021] Preferably, in step S3, for the first optimized sampling module, use 3×3 depth convolution operation and similarity calculation between feature maps to generate a class of compensation feature maps. This process ensures that the detailed features in the remote sensing image can be accurately captured and optimized by analyzing the relationship between pixels and color channels. By concatenating the compensation feature maps and the loss image, an optimized encoded image is generated, enabling the remote sensing mapping image enhancement model to fully retain local details and texture information during the remote sensing mapping image enhancement process, thereby improving the overall quality of the remote sensing mapping image

[0022] Preferably, in step S4, for the second optimized sampling module, obtain the single low-quality remote sensing mapping image I output by the initialization module, the single loss low-quality remote sensing mapping image I u , the loss image Iloss , I ∈ R h×w×3 , I u ∈ R h×w×3 , I loss ∈ R h×w×3 , where h, w, and 3 represent the height, width, and channels of a single low-quality remote sensing mapping image I, a single low-quality remote sensing mapping image I with losses, u , and a loss image I loss respectively, and R represents the real number space, and then the feature F TS1 and the feature F TS2 , F TS1 = Conv 3×3 (I), F TS2 = Conv 3×3 (I u ), where h1, w1, and c1 represent the height, width, and channels of the feature F TS1 and the feature F TS2 respectively, and then the weighted average feature value F TS1 and the feature F TS2 at each pixel position (i, j) is calculated, F TS1 (i, j) and F TS2 (i, j), where i ∈ [0, h1 - 1], j ∈ [0, w1 - 1], · represents dot product, denotes the feature value at each adjacent pixel position (i + m, j + n) in the feature F TS1 , denotes the feature value at each adjacent pixel position (i + m, j + n) in the feature F TS2 , ∑ is the summation symbol, w(m, n) represents the weight function, e is the natural logarithm, m represents the row offset relative to the central pixel, n represents the column offset relative to the central pixel, m 2 + m 2 represents the squared distance between the adjacent pixel and the central pixel, σ represents the standard deviation of the Gaussian function, w(m, n) is used to assign weights to the feature values at each adjacent pixel position (i + m, j + n), m ∈ (-2, 2), n ∈ (-2, 2), and then the second-class compensation feature map C2 is calculated,

[0023] C2 = TConv(β(F TS1 (i, j) - F TS2((i, j)), target size = h × w × 3), where β represents the adaptive hyperparameter for adjusting the compensation intensity, - represents element-wise subtraction, TConv represents transposed convolution, and target size = h × w × 3 represents that the output target dimension of the transposed convolution TConv is h × w × 3. Concatenate the secondary compensation feature map C2 and the loss image I loss in the channel dimension to obtain the mixed image I TSC , I TSC ∈R h×w×6 , and then obtain the optimized encoded image I CE2 ,

[0024] I CE2 = Down(DWConv 3×3 (Concat(Linear(I TSC ), I))), where Linear represents the linear layer, Concat represents concatenation in the channel dimension, DWConv 3×3 represents a 3×3 depth convolution, Down represents downsampling, and the optimized encoded image I CE2 is the output of the second optimized sampling module.

[0025] Preferably, in step S4, for the second optimized sampling module, by calculating the eigenvalue relationship between adjacent pixels and the central pixel, a secondary compensation feature map is generated. By assigning weights based on the spatial distance between pixels, the features of adjacent pixels in the image can be more closely combined. This processing method effectively improves the detail restoration ability of remote sensing mapping images, especially in key areas such as edges and textures. By concatenating the compensation feature map and the loss image, a more optimized encoded image is generated, further improving the accuracy of remote sensing mapping image enhancement.

[0026] Preferably, in step S5, for the feature processing module, a symmetric encoding structure is used, implemented by U-Net, which includes an input layer, an encoder, a bottleneck layer, a decoder, and an output layer. The dimensions of the input image of the input layer and the output image of the output layer are the same. Both the input layer and the output layer are implemented by 3×3 convolutions. The encoder includes multiple stacked convolutional layers, ReLU activation functions, and max pooling layers. The bottleneck layer is between the encoder and the decoder and is implemented using downsampling. The decoder includes transposed convolutional layers and skip connections.

[0027] Preferably, in step S6, for the remote sensing mapping image enhancement model, input a single low-quality remote sensing mapping image I into the initialization module, I ∈ R h×w×3 , where h, w, and 3 represent the height, width, and channels of the single low-quality remote sensing mapping image I, and R represents the real number space, that is, h and w take real values. The initialization module outputs a single low-quality remote sensing mapping image I and a single low-quality loss remote sensing mapping image Iu and the loss image I loss and the low-quality remote sensing mapping image I d , I u ∈R h×w×3 , I loss ∈R h×w×3 , Then, the single low-quality remote sensing mapping image I and the single loss low-quality remote sensing mapping image I u and the loss image I loss are input into the optimization sampling module one and the optimization sampling module two. The optimization sampling module one outputs the optimized encoded image I CE1 , and the optimization sampling module two outputs the optimized encoded image I CE2 , The optimized encoded image I CE1 and the optimized encoded image I CE2 and the low-quality remote sensing mapping image I d are concatenated in the channel dimension to obtain the multi-mixed image I MIX , The multi-mixed image I MIX is input into a 3×3 convolution to obtain the intermediate processed image I MID , The intermediate processed image I MID is input into the feature processing module to obtain the final optimized image I Fin , The final optimized image I Fin is input into upsampling to obtain a single high-quality remote sensing mapping image I + , I + ∈R h×w×3 , and the single high-quality remote sensing mapping image I + is the output of the remote sensing mapping image enhancement model.

[0028] Compared with the prior art, the present invention has the following technical effects:

[0029] The technical solution provided by the present invention proposes a remote sensing mapping image enhancement model. The initialization module provides various types of images for subsequent operations. The optimization sampling module one effectively enhances the detail performance and the accuracy of local features of the remote sensing image through convolution operations and the calculation of compensated feature maps. The optimization sampling module two further improves the correlation between adjacent pixels and the detail restoration effect in the remote sensing image by calculating the compensated feature map through weighted average. The remote sensing mapping image enhancement model significantly improves the detail restoration and overall quality of the low-quality remote sensing mapping image through the combined optimization of multiple modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the flowchart of the remote sensing mapping image enhancement process provided by the present invention.

[0031] Figure 2 It is the structural diagram of the remote sensing mapping image enhancement model provided by the present invention.

[0032] Figure 3 It is a single low-quality remote sensing mapping image provided by the present invention.

[0033] Figure 4 It is a single high-quality remote sensing mapping image provided by the present invention.

[0034] Figure 5 It is a partial single low-quality remote sensing mapping image provided by the present invention.

[0035] Figure 6 It is a partial single high-quality remote sensing mapping image provided by the present invention. Detailed implementation manners

[0036] The present invention aims to propose a method for enhancing remote sensing mapping images, and proposes a remote sensing mapping image enhancement model. Among them, the initialization module provides various types of images for subsequent operations. The first optimized sampling module effectively enhances the detail performance and the accuracy of local features of the remote sensing image through convolution operations and the calculation of compensated feature maps. The second optimized sampling module further improves the correlation between adjacent pixels and the detail restoration effect in the remote sensing image by calculating the compensated feature maps through weighted averaging. The remote sensing mapping image enhancement model significantly improves the detail restoration and overall quality of low-quality remote sensing mapping images through the combined optimization of multiple modules.

[0037] Please refer to Figure 1 shown in the following, a method for enhancing remote sensing mapping images in the embodiments of the present application:

[0038] S1. Production of the remote sensing mapping image data set: Collect 600 high-quality remote sensing mapping images and produce a remote sensing mapping image data set;

[0039] S2. Construct an initialization module, including 3-fold downsampling and 3-fold upsampling;

[0040] S3. Construct the first optimized sampling module, including 3×3 depth convolution and calculation of a type of compensated feature map;

[0041] S4. Construct the second optimized sampling module, including 3×3 depth convolution and calculation of a second type of compensated feature map;

[0042] S5. Construct a feature processing module, using a symmetric coding structure;

[0043] S6. Construct a remote sensing mapping image enhancement model, including input, initialization module, the first optimized sampling module, the second optimized sampling module, 3×3 convolution, feature processing module, upsampling, output;

[0044] S7. Training and use of remote sensing mapping image enhancement model: Use a remote sensing mapping image dataset to train a remote sensing mapping image enhancement model. After training, input a single low-quality remote sensing mapping image into the remote sensing mapping image enhancement model, and the remote sensing mapping image enhancement model outputs a single high-quality remote sensing mapping image.

[0045] Furthermore, in step S1, for the remote sensing mapping image dataset, collect 600 high-quality remote sensing mapping images. The height and width of each high-quality remote sensing mapping image are both 4800. Perform downsampling, transposed convolution operation, Gaussian blur, and motion blur on the 600 high-quality remote sensing mapping images to obtain 600 low-quality remote sensing mapping images. The height and width of each low-quality remote sensing mapping image are both 4800, and form 600 training pairs with the 600 high-quality remote sensing mapping images, thus forming a remote sensing mapping image dataset.

[0046] Furthermore, in step S2, for the initialization module, input a single low-quality remote sensing mapping image I, I ∈ R h ×w×3 , where h, w, and 3 represent the height, width, and channels of the single low-quality remote sensing mapping image I, and R represents the real number space, that is, h and w take real values. Downsample the single low-quality remote sensing mapping image I by 3 times to obtain a single 3-times downsampled low-quality remote sensing mapping image I d , which is one-third of h and w. Upsample the single 3-times downsampled low-quality remote sensing mapping image I d by 3 times to obtain a single loss low-quality remote sensing mapping image I u , I u ∈ R h×w×3 . Subtract the single loss low-quality remote sensing mapping image I u from the single low-quality remote sensing mapping image I element by element to obtain a loss image I loss , I loss ∈ R h×w×3 , I loss = I - I u , where - represents element-by-element subtraction. The single low-quality remote sensing mapping image I, the single loss low-quality remote sensing mapping image I u , the loss image I loss , and the low-quality remote sensing mapping image I d are the outputs of the initialization module.

[0047] Furthermore, in step S3, for the first optimization sampling module, obtain the single low-quality remote sensing mapping image I, the single loss low-quality remote sensing mapping image I u , and the loss image I loss output by the initialization module, I ∈ R h×w×3 , I u∈R h×w×3 , I loss ∈R h×w×3 , where h, w, and 3 represent the height, width, and channels of a single low-quality remote sensing mapping image I, a single lossy low-quality remote sensing mapping image I u , the lossy image I loss . R represents the real number space, and then the feature F OS1 and the feature F OS2 , F OS1 = DWConv 3×3 (I), F OS2 = DWConv 3×3 (I u ), DWConv 3×3 represents a 3×3 depth convolution, , where h1, w1, and c1 represent the height, width, and channels of the feature F OS1 and the feature F OS2 . Then, through the feature F OS1 and the feature F OS2 , for each pixel position (i, j) and each color channel c, the similarity Similar OS1 between the feature F OS2 and the feature F i，j，c is calculated,

[0048]

[0049] represents the feature F OS1 for each pixel position (i, j) and each color channel c, represents the feature F OS2 for each pixel position (i, j) and each color channel c, i ∈ [0, h1 - 1], j ∈ [0, w1 - 1], c ∈ [0, c1 - 1], · represents dot product, represents 's L2 norm, represents 's L2 norm. Then, a class of compensated feature maps C1, C1 ∈ R h×w×3 is calculated, where TConv represents transposed convolution, represents an adaptive hyperparameter for adjusting the compensation intensity, target size = h × w × 3 represents the output target dimension of the transposed convolution TConv as h × w × 3. The class of compensated feature maps C1 and the lossy image I loss are concatenated in the channel dimension to obtain the mixed image I OSC , I OSC ∈ R h×w×6 . Then, the optimized encoded image I CE1 is obtained,

[0050] I CE1 = Down(DWConv 3×3 (Concat(Linear(I OSC ), I))), where Linear represents the linear layer, Concat represents concatenation in the channel dimension, DWConv 3×3 represents a 3×3 depth convolution, Down represents downsampling, and the optimized encoded image I CE1 is the output of the first optimized sampling module.

[0051] Furthermore, in step S4, for the second optimized sampling module, a single low-quality remote sensing mapping image I obtained from the output of the initialization module, a single lossy low-quality remote sensing mapping image I u , the loss image I loss , I ∈ R h×w×3 , I u ∈ R h×w×3 , I loss ∈ R h×w×3 , where h, w, and 3 represent the height, width, and channels of the single low-quality remote sensing mapping image I, the single lossy low-quality remote sensing mapping image I u , and the loss image I loss , and R represents the real number space. Then the feature F TS1 and the feature F TS2 are obtained, where F TS1 = Conv 3×3 (I), F TS2 = Conv 3×3 (I u ), , where h1, w1, and c1 represent the height, width, and channels of the feature F TS1 and the feature F TS2 . Then the weighted average feature values F TS1 and F TS2 at each pixel position (i, j) are calculated, where i ∈ [0, h1 - 1], j ∈ [0, w1 - 1], TS1 (i, j) and F TS2 (i, j), where · represents the dot product, TS1 represents the feature value at each adjacent pixel position (i + m, j + n) in the feature F TS2 TS2 e is the natural logarithm, m represents the row offset relative to the central pixel, n represents the column offset relative to the central pixel, m 2 +n 2 represents the squared distance between the adjacent pixel and the central pixel, σ represents the standard deviation of the Gaussian function, w(m, n) is used to assign weights to the feature values at each adjacent pixel position (i + m, j + n), m ∈ (-2, 2), n ∈ (-2, 2), and then the second-class compensation feature map C2 is calculated.

[0052] C2 = TConv(β(F TS1 (i, j) - F TS2 (i, j)), target size = h × w × 3), where β represents the adaptive hyperparameter for adjusting the compensation intensity, - represents element-wise subtraction, TConv represents transposed convolution, target size = h × w × 3 represents that the output target dimension of the transposed convolution TConv is h × w × 3. The second-class compensation feature map C2 and the loss image I loss are concatenated in the channel dimension to obtain the mixed image I TSC I TSC ∈ R h×w×6 and then the optimized encoded image I CE2 is obtained.

[0053]

[0054] I CE2 = Down(DWConv 3×3 (Concat(Linear(I TSC ), I))), where Linear represents the linear layer, Concat represents concatenation in the channel dimension, DWConv 3×3 represents 3×3 depth convolution, Down represents downsampling, and the optimized encoded image I CE2 is the output of the second optimized sampling module.

[0055] Furthermore, in step S5, for the feature processing module, a symmetric encoding structure is used, implemented by U-Net, which includes an input layer, an encoder, a bottleneck layer, a decoder, and an output layer. The input image of the input layer and the output image of the output layer have the same dimension, and both the input layer and the output layer are implemented by 3×3 convolution. The encoder includes multiple stacked convolutional layers, ReLU activation functions, and max pooling layers. The bottleneck layer is between the encoder and the decoder and is implemented using downsampling. The decoder includes transposed convolutional layers and skip connections.

[0056] Furthermore, in step S6, for the remote sensing mapping image enhancement model, its structure is as Figure 2As shown, input a single low-quality remote sensing mapping image I into the initialization module, I ∈ R 4800×4800×3 , where 4800, 4800, and 3 represent the height, width, and channels of the single low-quality remote sensing mapping image I, and R represents the real number space. The initialization module outputs the single low-quality remote sensing mapping image I, the single loss low-quality remote sensing mapping image I u , the loss image I loss , the low-quality remote sensing mapping image I d , I u ∈R 4800×4800×3 , I loss ∈R 4800×4800×3 , I d ∈R 1600×1600×3 , and then input the single low-quality remote sensing mapping image I, the single loss low-quality remote sensing mapping image I u , the loss image I loss into the optimization sampling module one and the optimization sampling module two. The optimization sampling module one outputs the optimized encoded image I CE1 , I CE1 ∈R 1600×1600×3 , and the optimization sampling module two outputs the optimized encoded image I CE2 , I CE2 ∈R 1600×1600×3 . Concatenate the optimized encoded image I CE1 , the optimized encoded image I CE2 , and the low-quality remote sensing mapping image I d in the channel dimension to obtain the multi-mixed image I MIX , I MIX ∈R 1600×1600×9 . Input the multi-mixed image I MIX into a 3×3 convolution to obtain the intermediate processed image I MID , I MID ∈R 1600×1600×3 . Input the intermediate processed image I MID into the feature processing module to obtain the final optimized image I Fin , I FIN ∈R 1600×1600×3 . Input the final optimized image I Fin into upsampling to obtain a single high-quality remote sensing mapping image I + , I + ∈R 4800×4800×3 . The single high-quality remote sensing mapping image I + is the output of the remote sensing mapping image enhancement model.

[0057] Further, in step S7, Figure 3 As shown, it is a single low-quality remote sensing mapping image. The height and width of the single low-quality remote sensing mapping image are both 4800. Input the single low-quality remote sensing mapping image into the remote sensing mapping image enhancement model to obtainFigure 4 The single high-quality remote sensing mapping image shown has a height and width of 4800. Compared with a single low-quality remote sensing mapping image, the single high-quality remote sensing mapping image has better details and higher image quality. For the convenience of comparison, from Figure 3 the lower right corner area of the single low-quality remote sensing mapping image shown, Figure 5 a part of the single low-quality remote sensing mapping image shown is extracted. From Figure 4 the lower right corner area of the single high-quality remote sensing mapping image shown, Figure 6 a part of the single high-quality remote sensing mapping image shown is extracted. The height and width of the part of the single low-quality remote sensing mapping image and the part of the single high-quality remote sensing mapping image are both 200, and the displayed content is the same.

[0058] Furthermore, in step S7, the remote sensing mapping image enhancement model is trained using the Python language, the Pytorch framework, and an NVIDIA GPU 3090. The number of training rounds is 300, the number of batches in one round of training is 16, the Adam optimizer is used, and the initial learning rate is set to 0.0001. For 600 training pairs, 400 training pairs are used as the training set, 100 training pairs are used as the validation set, and 100 training pairs are used as the test set. The peak signal-to-noise ratio and structural similarity are used as the main evaluation criteria, and the subjective visual effect of the human eye is used as the auxiliary evaluation criteria. The training batch model with the highest peak signal-to-noise ratio, the highest structural similarity value, and the best subjective visual effect of the human eye is comprehensively selected as the final remote sensing mapping image enhancement model.

[0059] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A remote sensing imaging image enhancement processing method, characterized in that: The following steps are involved: S1. Remote sensing mapping image dataset preparation: collect multiple high-quality remote sensing mapping images and prepare a remote sensing mapping image dataset; S2, construct an initialization module, including 3 times downsampling and 3 times upsampling; the initialization module inputs a single low-quality remote sensing mapping image , , , and 3 represents the image height, width and passages, represents the real number space, that is and Take the value as a real number, Perform 3x downsampling to obtain a single 3x downsampled low-quality remote sensing mapping image , ,Will Upsampling by 3 times to obtain a single low-quality remote sensing image , ,use Single-shot low-quality remote sensing imagery Perform element-by-element subtraction to obtain the loss image , ,in ,in , , and is the output of the initialization module; S3, construct an optimized sampling module 1, including 3×3 depth convolution and calculation of a type of compensation feature map; the optimized sampling module 1 obtains the output of the initialization module and , , then get the features and Features ,in , , represents a 3×3 depthwise convolution, , , , and Representative features and Features The height, width and channels of the pixel are then The eigenvalue corresponding to the color channel c and Calculate features and Features Similarity between , , represent The L2 norm of represent The L2 norm of , and then calculate a class of compensation feature maps , , ,in represents transposed convolution, represents the adaptive hyperparameter that adjusts the compensation strength, Represents transposed convolution The output target dimension is , a class of compensation feature maps and loss image Stitching is performed in the channel dimension to obtain a mixed image , , and then get the optimized encoded image , , ,in represents a linear layer, Represents splicing in the channel dimension, represents a 3×3 depthwise convolution, Represents downsampling, optimizing the encoded image is the output of the optimized sampling module 1; S4, constructing an optimized sampling module 2, including 3×3 depth convolution and calculating two types of compensation feature maps; the optimized sampling module 2 obtains the output of the initialization module , , , and then get the features and Features , , , , , then calculate the features and Features Each pixel position The weighted average eigenvalue of and , , , Indicated in the feature Each adjacent pixel position The characteristic value of Indicated in the feature Each adjacent pixel position The characteristic value of is the summation symbol, represents the weight function, , is the natural logarithm, Represents the row offset relative to the center pixel, represents the column offset relative to the center pixel, represents the standard deviation of the Gaussian function, For each adjacent pixel position The eigenvalues ​​of are assigned weights, , , and then calculate the second type of compensation feature map , , ,in represents the adaptive hyperparameter that adjusts the compensation strength, Represents transposed convolution The output target dimension is , the second type of compensation feature map and loss image Stitching is performed in the channel dimension to obtain a mixed image , , and then get the optimized encoded image , , ,in is the output of the optimized sampling module 2; S5. Construct a feature processing module using a symmetric coding structure; the feature processing module uses a symmetric coding structure and is implemented by U-Net, comprising an input layer, an encoder, a bottleneck layer, a decoder, and an output layer, wherein the input image of the input layer and the output image of the output layer have the same dimension, the input layer and the output layer are both implemented by 3×3 convolution, the encoder comprises multiple stacked convolution layers, ReLU activation functions, and maximum pooling layers, the bottleneck layer is between the encoder and the decoder, and is implemented by downsampling, and the decoder comprises a transposed convolution layer and a skip connection; S6, construct a remote sensing image enhancement model, including input, initialization module, optimization sampling module 1, optimization sampling module 2, 3×3 convolution, feature processing module, upsampling, output; the remote sensing image enhancement model, input To the initialization module, the initialization module outputs a single low-quality remote sensing image , , , , then , , Input to the optimized sampling module 1 and the optimized sampling module 2, the optimized sampling module 1 outputs the optimized coded image , the optimized sampling module 2 outputs the optimized coded image ,Will , , Multiple mixed images are obtained by splicing in the channel dimension , ,Will Input to 3×3 convolution to get the intermediate processed image , ,Will Input into the feature processing module to obtain the final optimized image , ,Will Input to upsample to get a single high-quality remote sensing image , ; S7. Training and use of remote sensing image enhancement model: Use the remote sensing image dataset to train the remote sensing image enhancement model. After the training is completed, input a single low-quality remote sensing image to the remote sensing image enhancement model, and the remote sensing image enhancement model outputs a single high-quality remote sensing image.

2. The remote sensing image enhancement processing method according to claim 1, characterized in that: The remote sensing mapping image data set collects multiple high-quality remote sensing mapping images, performs downsampling and transposed convolution operations, Gaussian blur, and motion blur on the multiple high-quality remote sensing mapping images, obtains multiple low-quality remote sensing mapping images, and forms multiple training pairs with the multiple high-quality remote sensing mapping images, thereby forming a remote sensing mapping image data set.

Images