Remote sensing image hue information maintaining method based on diffusion super-division model

Abstract

The invention discloses a remote sensing image hue information keeping method based on a diffusion super-division model, and the method comprises the steps: constructing a data set, carrying out the image preprocessing of an original high-resolution image, obtaining a feature image after preprocessing, and dividing the feature image into a test set and a verification set; constructing a diffusion super-division model; designing a residual function of the diffusion super-division model; performing cross reference on the training set and the verification set, combining a residual function, and optimizing model parameters by continuously minimizing the residual function to train a diffusion super-division model to obtain a trained diffusion super-division model; deploying the trained diffusion super-division model on a remote sensing satellite to form a hue information keeping module based on the diffusion super-division model; a remote sensing satellite generates an original image needing to be processed, the original image is input into a hue information keeping module based on a diffusion super-division model, model reasoning is conducted through the trained diffusion super-division model, a high-resolution image is obtained from a low-resolution image in a super-division mode, and image details are restored on the basis of keeping hue information.

Description

Technical Field

[0001] The present invention relates to a method for preserving the tone information of a remote sensing image based on a diffusion super-resolution model, which belongs to the field of remote sensing image processing and is particularly suitable for processing low-resolution remote sensing images using artificial intelligence technology, automatically increasing the texture information of the remote sensing image and keeping the tone information unchanged, thereby ultimately achieving the purpose of improving the image clarity. Background Art

[0002] Image super-resolution technology is widely used in the field of remote sensing, and a large number of algorithms have been formed in long-term research. From the early interpolation-based methods, reconstruction-based methods, and learning-based methods, we have turned to efficient methods to predict the high-frequency information missing in low-resolution images. Traditional super-resolution methods for remote sensing images can be roughly divided into two types: one is the traditional signal filtering method, and the second is the deep learning method that uses a graphics computing platform to train neural networks.

[0003] The traditional signal filtering-based approach includes a fast and practical single image super-resolution method that uses a lookup table for acceleration. This method trains a deep SR network with a small receptive field and then transfers the output values ​​of the learned deep model to a lookup table. A comparative study between super-resolution methods based on polynomial interpolation and Whittaker filter interpolation analyzes and compares super-resolution techniques using polynomial and Whittaker filter interpolation. It mainly focuses on image resolution improvement in computer vision, especially single-frame and multi-frame super-resolution. Although they are simple and fast, some high-frequency information is destroyed in the process. Learning-based methods include multiple linear mapping anchored neighbor regression, which improves the quality of high-resolution images obtained from a single low-resolution input by learning a series of mapping relationships between low-resolution and high-resolution feature subspaces. Its innovation lies in the use of multiple linear mappings combined with anchored neighbor regression to transform low-resolution feature subspaces into high-resolution subspaces. Learning-based methods rely on a large number of low-resolution and high-resolution images as prior information. Reconstruction-based methods include using multiple mixed prior models to solve the problem of reconstructing a high-resolution image from a single low-resolution image. The innovation is to use a selective patch processing method based on differential curvature to analyze the image, and then use a hybrid prior model to group these patches for learning, thereby improving the effect of super-resolution reconstruction. However, it has the disadvantages of slow convergence and high computational cost.

[0004] With the rapid development of artificial intelligence and the continuous strengthening of hardware computing platforms, deep learning has become a powerful tool for solving computer vision problems. Deep learning has profoundly promoted the development of remote sensing, including land cover mapping, environmental parameter retrieval, data fusion and downscaling, and information reconstruction and prediction. Along with the variation and development of various neural networks, frameworks such as convolutional neural networks (CNN), generative adversarial networks (GAN), transformer models, and diffusion models have achieved many results in the field of image super-resolution.

[0005] The denoising diffusion probability model is a new image generation method based on the diffusion probability model. The diffusion model reverses the noise diffusion process by training a parameterized Markov chain. The structure of the diffusion model mainly includes forward noise image generation, and reverse noise prediction and removal. When adding noise in the forward direction, a longer step value needs to be set, and each step adds a certain amount of random noise to the image of the previous step. When the step length is long enough, the original image becomes a random noise image that completely conforms to the normal distribution. The input of the reverse process is a random noise image, and a clear image is obtained after continuously removing the noise.

[0006] The famous Stable Diffusion algorithm developed a high-resolution image synthesis technique based on a latent diffusion model, such as Figure 1 As shown in the figure, a cross-attention layer is used to improve the quality and relevance of the generated image, and text or other modal data is used as input to act as a conditional control. Experiments have shown that the algorithm performs well in tasks such as image generation, editing, and super-resolution. In addition to image generation, the diffusion model can also be used in image segmentation, image denoising, image restoration, image super-resolution, image editing, and other fields. SR3 is an image super-resolution method based on a denoising diffusion probability model. SR3 is able to generate images that are both realistic and faithful to low-resolution inputs. It starts with pure Gaussian noise and gradually refines the noisy output through an iterative denoising process. The output of SR3 on the ImageNet super-resolution task scores high on FID and IS, but low on PSNR and SSIM.

[0007] Compared with traditional super-resolution algorithms, the diffusion model can not only generate high-quality images, but also has a relatively stable training process and the model error is easy to converge. However, during the training process, it was found that due to the introduction of random noise, the remote sensing super-resolution images generated by the diffusion model will have hue distortion.

[0008] When training a traditional diffusion model, noise is first added to the original image. The added noise is known and related to the step size. During training, a deep network model is used to predict the noise. The loss function is defined as the difference between the predicted noise and the added noise. The purpose is to make the noise output by the deep model as consistent as possible with the simulated noise. During inference, when the step size of the diffusion model is large, the original image can be restored by continuously predicting and removing noise. However, since there is no constraint on the image hue during training, it is easy for the hue information to be inconsistent with the original image during inference. Summary of the invention

[0009] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art and provide a method for preserving the hue information of remote sensing images based on a diffusion super-resolution model, which can solve the hue distortion problem of the diffusion model during super-resolution of remote sensing images while restoring image details.

[0010] The technical solution of the present invention is: a method for preserving hue information of remote sensing images based on a diffusion super-resolution model, comprising:

[0011] Constructing a dataset: preprocessing the original high-resolution images to obtain preprocessed feature images, and dividing them into a test set and a validation set;

[0012] Construct a diffusion super-resolution model;

[0013] Design the residual function of the diffusion super-resolution model;

[0014] Cross-reference the training set and the validation set, and combine the residual function to train the diffusion super-resolution model by continuously minimizing the residual function to optimize the model parameters, and obtain a trained diffusion super-resolution model;

[0015] Deploy the trained diffusion super-resolution model to the remote sensing satellite to form a hue information preservation module based on the diffusion super-resolution model;

[0016] The remote sensing satellite generates the original image that needs to be processed and inputs it into the tone information preservation module based on the diffusion super-resolution model. The trained diffusion super-resolution model is used for model inference to super-resolve high-resolution images from low-resolution images and restore image details while maintaining the tone information.

[0017] Preferably, the image source for constructing the data set is: high-resolution satellite multispectral images taken during a random year.

[0018] Preferably, the image preprocessing operation includes:

[0019] The original high-resolution satellite multispectral image is superimposed with noise, which is recorded as the first superimposed image;

[0020] Downsampling the original high-resolution satellite multispectral image to obtain a low-resolution image, then upsampling the low-resolution image to a high-resolution image size, and recording the upsampled image as a second stacked image;

[0021] The first superimposed image and the second superimposed image are combined to obtain a feature image with twice the number of channels as the preprocessed feature image.

[0022] Preferably, the construction of the diffusion super-resolution model includes: four ResNetBlock modules, three downsampling modules, three ConvNeXtBlock modules, and three upsampling modules, which together constitute a U-Net network structure; specifically:

[0023] The three groups of ResNetBlock modules and the three downsampling modules correspond to each other to form three groups of encoding modules, which are respectively recorded as the first encoding module, the second encoding module, and the third encoding module. The three groups of encoding modules are connected in sequence; each group of encoding modules is used to solve the degradation problem in the network and smoothly transmit the knowledge learned by the network in the early stage to the later stage of the network;

[0024] The three ConvNeXtBlock modules and the three upsampling modules correspond to each other to form three groups of decoding modules, which are respectively recorded as the first decoding module, the second decoding module, and the third decoding module; the three groups of decoding modules are connected in sequence; each group of decoding modules is used to gradually refine the result of super-resolution reconstruction, and the model can fuse features at different scales, thereby further improving the image quality on the basis of the previous group;

[0025] The fourth ResNetBlock module is connected to the third encoding module as a skip connection layer, with its input connected to three groups of encoding modules and its output connected to three groups of decoding modules, which are used to enhance gradient flow and improve the stability of network training.

[0026] Preferably, the preprocessed feature image is input into the diffusion super-resolution model, and the image is sequentially passed through a first encoding module, a second encoding module, and a third encoding module, and an encoded noisy low-resolution image is output;

[0027] The noisy low-resolution image output by the encoding module passes through the skip connection layer and outputs the predicted noise image;

[0028] The predicted noise image output by the skip connection layer passes through the first decoding module, the second decoding module, and the third decoding module in sequence to output a predicted high-resolution image.

[0029] Preferably, the residual function Loss is:

[0030]

[0031] L1 and L2 are loss values, among which L1 is less sensitive to outliers and provides robustness; L2 is more sensitive to small deviations in the data and provides accuracy; SSIM is the structural similarity index, which ensures the structural fidelity of the image.

[0032] Preferably, the loss value L1 and the loss value L2 are respectively:

[0033]

[0034] Among them, x0 represents the original high-resolution image, Represents the predicted high-resolution image.

[0035] Preferably, the structural similarity index SSIM is:

[0036]

[0037] Where: μ x , μ y They are image x0 and The mean of represents the average brightness; They are image x0 and The standard deviation of , indicating the contrast of the image; σ xy represents the image x0 and The covariance between them measures the structural similarity between images; C1 and C2 are small constants used to maintain the stability of the calculation;

[0038] x0 represents the original high-resolution image, Represents the predicted high-resolution image.

[0039] Preferably, when training the diffusion super-resolution model:

[0040] The predicted high-resolution image output by the diffusion super-resolution model and the original high-resolution image are brought into the residual function to calculate the residual function loss value. The back propagation algorithm is used to modulate the model parameters according to the residual function loss value to make the loss value smaller and smaller, thereby realizing the model training.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] (1) The present invention achieves the effect of generating high-quality images by proposing a diffusion model, and the training process is relatively stable, and the model error is easy to converge;

[0043] (2) The present invention proposes a diffusion improvement model to achieve the effect of restoring image details while solving the hue distortion problem of the diffusion model during remote sensing image super-resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1Schematic diagram of the principle of the diffusion model in the background technology of the present invention: from left to right is noise addition, and from right to left is denoising;

[0045] Figure 2 Schematic diagram of the super-resolution model structure of the present invention;

[0046] Figure 3 It is a schematic diagram of the diffusion super-resolution model of the present invention;

[0047] Figure 4 It is the ResNetBlock structure diagram of the present invention;

[0048] Figure 5 This is the ConvNexBlock structure diagram of the present invention. DETAILED DESCRIPTION

[0049] The model structure of the present invention is based on the diffusion model and improves the standard diffusion model. The specific improvements include two points: first, a step of predicting a high-resolution image is added, that is, based on the predicted noise image, the image restoration formula is used to directly predict and restore the noisy remote sensing image to the original image; second, a tone-preserving loss function is designed to keep the predicted high-resolution image consistent with the tone of the original image. The structure of the super-resolution model proposed by the present invention is shown in the figure below: Figure 2 shown.

[0050] The technical solution of the present invention mainly includes the construction of a diffusion super-resolution model and the design of a residual function. Unlike the standard diffusion model, the input of the image super-resolution algorithm is a low-resolution image, so when designing the model, the low-resolution image can be used as a constraint condition and combined with a random noise map as input.

[0051] The present invention provides a method for preserving hue information of remote sensing images based on a diffusion super-resolution model, and the specific steps are as follows:

[0052] S1. Construct a data set, preprocess the original high-resolution image, obtain the preprocessed feature image, and divide it into a test set and a validation set;

[0053] S2, build a diffusion super-resolution model;

[0054] S3, design the residual function of the diffusion super-resolution model;

[0055] S4, cross-reference the training set and the validation set, and combine the residual function to train the diffusion super-resolution model by continuously minimizing the residual function to optimize the model parameters, so as to obtain a trained diffusion super-resolution model;

[0056] S5, deploying the trained diffusion super-resolution model to the remote sensing satellite to form a tone information preservation module based on the diffusion super-resolution model;

[0057] S6. The remote sensing satellite generates the original image to be processed and inputs it into the tone information preservation module based on the diffusion super-resolution model. The trained diffusion super-resolution model is used for model inference to super-resolve the high-resolution image from the low-resolution image and restore the image details while maintaining the tone information.

[0058] Specific:

[0059] 1. When constructing the dataset, the image sources are: multispectral images taken by high-resolution satellites during a random year;

[0060] When making a dataset:

[0061] Image preprocessing was performed on multispectral image patches of high-resolution satellites with cropped set sizes. 50,000 images with image sizes of 256×256 pixels and 64×64 pixels each constituted the training set, and 5,000 images with image sizes of 256×256 pixels and 64×64 pixels each constituted the training set;

[0062] The image preprocessing operations include:

[0063] The original high-resolution satellite multispectral image is superimposed with noise, which is recorded as the first superimposed image;

[0064] Downsampling the original high-resolution satellite multispectral image to obtain a low-resolution image, then upsampling the low-resolution image to a high-resolution image size, and recording the upsampled image as a second stacked image;

[0065] The first superimposed image and the second superimposed image are combined to obtain a feature image with twice the number of channels as the preprocessed feature image.

[0066] 2. Model construction

[0067] The construction of the diffusion super-resolution model includes: four ResNetBlock modules, three downsampling modules, three ConvNeXtBlock modules, and three upsampling modules, which together constitute the U-Net network structure, such as Figure 3 As shown; specifically:

[0068] A ResNetBlock module is connected to a downsampling module to form a group of encoding modules. Three groups of ResNetBlock modules and three downsampling modules form three groups of encoding modules in the above manner, which are respectively recorded as the first encoding module, the second encoding module, and the third encoding module. The above three groups of encoding modules are connected in sequence; each group of encoding modules is used to solve the degradation problem in the network and smoothly transmit the knowledge learned by the network in the early stage to the later stage of the network;

[0069] The fourth ResNetBlock module is connected to the third encoding module as a skip connection layer to enhance gradient flow and improve the stability of network training;

[0070] Three ConvNeXtBlock modules and three upsampling modules correspond one to one. One ConvNeXtBlock module and one upsampling module form a group of decoding modules. The three groups of decoding modules are respectively recorded as the first decoding module, the second decoding module, and the third decoding module. The above three groups of decoding modules are connected in sequence; each group of decoding modules is used to gradually refine the result of super-resolution reconstruction. The model can fuse features at different scales, thereby further improving the image quality on the basis of the previous group;

[0071] The input end of the skip connection layer is connected to three sets of encoding modules, and the output end is connected to three sets of decoding modules;

[0072] The preprocessed feature image is input into the diffusion super-resolution model, and the image passes through the first encoding module, the second encoding module, and the third encoding module in sequence, and the encoded noisy low-resolution image is output; the noisy low-resolution image output by the encoding module passes through the skip connection layer, and the predicted noise image is output; the predicted noise image output by the skip connection layer passes through the first decoding module, the second decoding module, and the third decoding module in sequence, and the predicted high-resolution image is output;

[0073] Each ResNetBlock module is used for image feature extraction, including two convolutional layers and a skip connection layer. The connection relationship of each structure is as follows: Figure 4 As shown;

[0074] Each ConvNeXtBlock module is used for multi-scale feature extraction of images, including three convolutional layers and one skip connection layer. The connection relationship of each structure is as follows: Figure 5 As shown;

[0075] The noise prediction model of the present invention adopts a U-Net network structure mixed with ResNetBlock and ConvNeXtBlock. It is mainly divided into a downsampling module, an intermediate module and an upsampling module. According to the needs of super-resolution from low resolution to high resolution, a self-attention mechanism is added when the feature map size is low resolution to retain higher dimensional information and improve the model's ability to predict noise. ResNetBlock is used to solve the degradation problem in the deep network, allowing the knowledge obtained by the network in the early stage of learning to be smoothly transmitted to the later stage of the network. Combined with ConvNeXtBlock, it can better adapt to features of different scales. ConvNeXtBlock has higher computational efficiency than ResNetBlock. ConvNeXt also provides better operator performance than ResNet, especially on large-scale data sets. Different combinations are selected according to the different sizes of feature maps during the downsampling process. Three ConvNeXtBlocks are used as a loop method in the upsampling process to process deep network and shallow network features through residual connections.

[0076] 3. Residual function

[0077] Compared with the residual function of the traditional diffusion model, which only considers the noise, the present invention also considers the difference between the model predicted image and the original image. The residual function is also improved, and a combination scheme is adopted:

[0078]

[0079] Among them, x0 represents the original image, represents a high-resolution image predicted by a noise prediction deep neural network;

[0080] The SSIM calculation formula is as follows:

[0081]

[0082] Where: μ x , μ y They are image x0 and The mean of represents the average brightness; They are image x0 and The standard deviation of , indicating the contrast of the image; σ xy represents the image x0 and The covariance between them measures the structural similarity between images; C1 and C2 are small constants used to maintain the stability of the calculation.

[0083]

[0084] The improved loss function enhances the balance between robustness and accuracy. L1 loss is less sensitive to outliers and provides robustness; L2 loss is more sensitive to small deviations in the data and provides accuracy. The stability of model training optimization is improved. Combining the two can provide smoother and more stable gradients during training, which helps to better optimize the model. SSIM is the structural similarity index, which ensures the structural fidelity of the image, especially in maintaining the edges and textures to ensure the integrity of the image structure.

[0085] 4. Training the Diffusion Super-resolution Model

[0086] The noisy image after noise addition is combined with the low-resolution image to obtain a feature image with double the number of channels. The feature image is encoded and decoded to obtain a noise prediction image. The noise prediction image and the noisy image are operated according to the diffusion model to obtain a predicted high-resolution image. The predicted high-resolution image and the original high-resolution image are input into the loss function model to obtain the loss value. The back propagation algorithm is used to modulate the model parameters according to the loss value to make the loss value smaller and smaller, thereby realizing the training of the model.

[0087] 5. Image description:

[0088] Original high-resolution image x0: an image cropped directly from a remote sensing satellite image, maintaining the original image details and resolution;

[0089] Low-resolution remote sensing image y0: The image obtained by scaling the original high-resolution image and then upsampling it.

[0090] Random Noise z: A randomly generated noise image with the same size as the image.

[0091] Noised remote sensing image: A noisy image obtained by adding random noise of a certain weight to the original high-resolution image using a diffusion loading model;

[0092] Predicted high-resolution image The image is calculated based on the predicted noise image, the noisy remote sensing image, and the loading step value.

[0093] Noise loading mathematical model:

[0094]

[0095] Using a linear uniform noise scheme, at each time value t, according to the predetermined noise level (given by β t series definition), add Gaussian noise to the data, and predict the image Mathematical model:

[0096]

[0097] Among them, x0 represents the original high-resolution image, z represents random noise, represents the predicted high-resolution image, represents the predicted noise image, β t represents the noise variance corresponding to time t.

[0098] Example:

[0099] 1) Training sample preparation

[0100] The data used in the experiment are multispectral data from high-resolution satellites with an original resolution of 0.5 meters, which are uniformly downsampled to 2 meters to obtain low-resolution images. When making super-resolution training samples, image blocks of a set size are randomly cropped from remote sensing images taken during a year as high-resolution data sets. After data preprocessing, 50,000 images of the training set with a size of 256×256 pixels and 64×64 pixels are obtained, and 5,000 images of the training set and validation set with a size of 256×256 pixels and 64×64 pixels are obtained. The 64x64 image is the result of downsampling the 256x256 image and is used as a low-resolution image.

[0101] 2) Forward calculation

[0102] The training samples include high-resolution images and corresponding low-resolution images. In this example, the dimension of the high-resolution image x0 is 3x256x256, and the dimension of the low-resolution image y0 is 3x64x64, where 3 is the number of channels. First, the low-resolution image is resampled to the size of the high-resolution image, that is, 3x64x64 is resampled to 3x256x256. Random noise z is added to the high-resolution image to obtain the noise image x t , and combined with the low-resolution image to obtain a feature image with 6 channels and a dimension of 6x256x256. First, the input image is encoded. After three residual networks ResnetBlock and downsampling operations, the feature image becomes 1024x32x32. Then 1024x32x32 is decoded to 3x256x256. The decoding includes 3 ConvNeXth and upsampling operations. The decoded result is used as the noise prediction value.

[0103] 3) Model optimization

[0104] The goal of the present invention during training is to make the model predict high-resolution images Same as the original high-resolution image x0.

[0105] Since the output of the model is a noisy image, the high-resolution image is first predicted based on the output of the model and the residual value is calculated.

[0106] Then the model is optimized based on the residual value. The optimization method used is stochastic gradient descent. Based on chain derivatives, the residual value is used to calculate the derivative of each unknown parameter, and then the unknown number is iteratively updated according to the gradient descent method. Each training sample is input into the model for optimization, and after a large amount of training, the optimized model parameters can be obtained.

[0107] 4) Model Reasoning

[0108] After the model training is completed, in order to super-resolve high-resolution images from low-resolution images, we need model reasoning; input the low-resolution image that needs to be super-resolved. In this example, the dimension of the low-resolution image is 3x64x64, and the low-resolution image is resampled to 3x256x256. Generate a random noise image z, and perform feature fusion with the resampled low-resolution image to obtain a 6x256x256 feature image. According to the maximum step value of the diffusion model (1000 in this embodiment), denoise the feature image. Input 1000, 999, ..., 1 in sequence. Each time a step value is input, a noise image is predicted, and the current input image is denoised based on the noise image. The image after the last denoising is used as the input for the next denoising. After 1000 cycles, the output image is the final super-resolved image.

[0109] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

Claims

1. A method for preserving hue information of remote sensing images based on a diffusion super-resolution model, characterized in that include: Constructing a dataset: preprocessing the original high-resolution images to obtain preprocessed feature images, and dividing them into a test set and a validation set; Construct a diffusion super-resolution model; Design the residual function of the diffusion super-resolution model; Cross-reference the training set and the validation set, and combine the residual function to train the diffusion super-resolution model by continuously minimizing the residual function to optimize the model parameters, and obtain a trained diffusion super-resolution model; Deploy the trained diffusion super-resolution model to the remote sensing satellite to form a hue information preservation module based on the diffusion super-resolution model; The remote sensing satellite generates the original image that needs to be processed and inputs it into the tone information preservation module based on the diffusion super-resolution model. The trained diffusion super-resolution model is used for model inference to super-resolve high-resolution images from low-resolution images and restore image details while maintaining the tone information.

2. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 1, characterized in that: The image sources for constructing the dataset are: high-resolution satellite multispectral images taken during a random one-year period.

3. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 2, characterized in that: Image preprocessing operations include: The original high-resolution satellite multispectral image is superimposed with noise, which is recorded as the first superimposed image; Downsampling the original high-resolution satellite multispectral image to obtain a low-resolution image, then upsampling the low-resolution image to a high-resolution image size, and recording the upsampled image as a second stacked image; The first superimposed image and the second superimposed image are combined to obtain a feature image with twice the number of channels as the preprocessed feature image.

4. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 1, characterized in that: The construction of the diffusion super-resolution model includes: four ResNetBlock modules, three downsampling modules, three ConvNeXtBlock modules, and three upsampling modules, which together constitute the U-Net network structure; specifically: The three groups of ResNetBlock modules and the three downsampling modules correspond to each other to form three groups of encoding modules, which are respectively recorded as the first encoding module, the second encoding module, and the third encoding module. The three groups of encoding modules are connected in sequence; each group of encoding modules is used to solve the degradation problem in the network and smoothly transmit the knowledge learned by the network in the early stage to the later stage of the network; The three ConvNeXtBlock modules and the three upsampling modules correspond to each other to form three groups of decoding modules, which are respectively recorded as the first decoding module, the second decoding module, and the third decoding module; the three groups of decoding modules are connected in sequence; each group of decoding modules is used to gradually refine the result of super-resolution reconstruction, and the model can fuse features at different scales, thereby further improving the image quality on the basis of the previous group; The fourth ResNetBlock module is connected to the third encoding module as a skip connection layer, with its input connected to three groups of encoding modules and its output connected to three groups of decoding modules, which are used to enhance gradient flow and improve the stability of network training.

5. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 4, characterized in that: Inputting the preprocessed feature image into the diffusion super-resolution model, the image passes through the first encoding module, the second encoding module, and the third encoding module in sequence, and outputting the encoded noisy low-resolution image; The noisy low-resolution image output by the encoding module passes through the skip connection layer and outputs the predicted noise image; The predicted noise image output by the skip connection layer passes through the first decoding module, the second decoding module, and the third decoding module in sequence to output a predicted high-resolution image.

6. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 1, characterized in that: The residual function Loss is: L1 and L2 are loss values, among which L1 is less sensitive to outliers and provides robustness; L2 is more sensitive to small deviations in the data and provides accuracy; SSIM is the structural similarity index, which ensures the structural fidelity of the image.

7. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 6, characterized in that: The loss values ​​L1 and L2 are: Among them, x0 represents the original high-resolution image, Represents the predicted high-resolution image.

8. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 6, characterized in that: The structural similarity index SSIM is: Where: μ x , μ y They are image x0 and The mean of represents the average brightness; They are image x0 and The standard deviation of , indicating the contrast of the image; σ xy represents the image x0 and The covariance between them measures the structural similarity between images; C1 and C2 are small constants used to maintain the stability of the calculation; x0 represents the original high-resolution image, Represents the predicted high-resolution image.

9. The method for preserving hue information of remote sensing images based on a diffusion super-resolution model according to claim 1, characterized in that: When training the diffusion super-resolution model: The predicted high-resolution image output by the diffusion super-resolution model and the original high-resolution image are brought into the residual function to calculate the residual function loss value. The back propagation algorithm is used to modulate the model parameters according to the residual function loss value to make the loss value smaller and smaller, thereby realizing the model training.

Images