Remote sensing image defogging method based on diffusion model

Abstract

The invention belongs to the technical field of computer vision and image processing, and particularly relates to a remote sensing image defogging method based on a diffusion model, which adopts a double-stage defogging network framework and comprises a first-stage multi-scale feature selection fusion network and a second-stage perceptual decoding residual diffusion model. The multi-scale feature selection fusion network processes an input image to generate a prediction image so as to realize a rough defogging effect, and the multi-scale feature selection fusion network in the first stage is based on a U-net architecture and fuses a multi-scale feature extraction module and a selection fusion module; the multi-scale feature extraction module is used for extracting features of different scales, and then the selective fusion module performs selective weighting and fusion on the features; in the second stage, the preprocessed image is then used as condition input of a residual diffusion model with perceptual decoding, a perceptual decoder optimizes residual estimation by further decoupling conditions, and the method improves the generation quality and effectively enhances the defogging performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and image processing, and particularly relates to a remote sensing image defogging method based on a diffusion model. BACKGROUND

[0002] Remote sensing images have important applications in environmental monitoring, disaster assessment and other tasks. However, the suspended particles such as fog, haze and clouds in the atmosphere can cause the image contrast to decrease and the details to be blurred, which affects the subsequent information extraction. Traditional defogging methods mostly rely on atmospheric scattering models or empirical priors, but they have limited performance in complex or non-uniform fog distribution scenarios. With the development of deep learning, convolutional neural networks (CNN) have made preliminary breakthroughs in defogging tasks by extracting image features. However, CNNs are limited by local receptive fields and are difficult to capture long-range dependencies, which limits their ability to model global structures. To make up for this deficiency, the Transformer architecture has been introduced into the defogging task, enhancing the global modeling capability, but its high computational cost is not conducive to practical deployment.

[0003] In recent years, generative models, especially diffusion models, have become a research hotspot in the field of image restoration by modeling the image latent distribution. The representative method RSHazeDiff combines Fourier transform and multi-stage training strategy, significantly improving the defogging performance of remote sensing images. However, its inference process relies on a large number of sampling steps, which is inefficient and limits its application scenarios. To solve this problem, the residual denoising diffusion model (RDDM) proposes a double diffusion mechanism, which decomposes the task into residual and noise sub-processes, improving the generation efficiency while ensuring the diversity of the generated images. Although RDDM performs well in multiple scenarios, it still has problems in restoring heavy fog images. Among the two existing diffusion models for defogging, one has high computational resources and long sampling time, and the other has short sampling time but poor image restoration effect. The existing technology does not balance the computational resources and image restoration effect well, and it is necessary to balance the computational cost and actual effect. SUMMARY

[0004] The purpose of the present application is to provide a remote sensing image defogging method based on a diffusion model, which solves the technical problem that the existing technology does not balance the computational resources and image restoration effect well, resulting in the existing technology either consuming computational resources or taking too long, or having poor image restoration effect.

[0005] The remote sensing image defogging method based on the diffusion model adopts a two-stage defogging network framework, the two-stage defogging network framework comprising a first-stage multi-scale feature selection fusion network MCESS-Unet and a second-stage perception decoding residual diffusion model; the multi-scale feature selection fusion network MCESS-Unet processes an input image to generate a predicted image to achieve a rough defogging effect, the first-stage multi-scale feature selection fusion network MCESS-Unet is based on a U-net architecture and fuses a multi-scale feature extraction module and a selection fusion module; the multi-scale feature extraction module is used to extract features of different scales, and the selection fusion module then selectively weights and fuses the features; in the second stage, the predicted image is then used as a conditional input of the perception decoding residual diffusion model, is connected with the input image and is input into the perception decoding residual diffusion model, and the perception decoding residual diffusion model optimizes residual estimation by further decoupling conditions and gradually denoises and restores the image.

[0006] Preferably, in the multi-scale feature selection fusion network MCESS-Unet, each encoder and decoder comprises a multi-scale feature extraction module, the final output of the encoder part is output to the decoder part after being processed by a feature fusion module, and a selection fusion module is arranged on each jump connection; the encoder extracts hierarchical features through the multi-scale feature extraction module and reduces the spatial dimension through maximum pooling, in the decoder, the spatial resolution is gradually reconstructed using standard convolution, and the selection fusion module is used to fuse the hierarchical features with the corresponding encoder features to replace the traditional jump connection.

[0007] Preferably, the multi-scale feature extraction module uses different sizes of convolution kernels and maximum pooling to capture features of different scales, integrates the output after tensor splicing through 1x1 convolution, and uses a channel attention mechanism to selectively enhance important features.

[0008] Preferably, the two input features of the selection fusion module are first fused with global context information through an efficient gating mechanism, and then adaptively adjusted using a channel attention mechanism to highlight key channel features; the two input features are then processed by a selection mechanism, and channel compression and excitation and spatial compression and excitation are applied; the outputs of the two are added to the residual, and an importance maximum value operator is used to select the feature value with the largest weight to output.

[0009] Preferably, the channel compression and excitation extracts spatial information through global average pooling, and then generates channel attention weights through a fully connected layer and an S-shaped function in sequence; the spatial compression and excitation applies an activation function after 1x1 convolution to enhance the feature map, create a spatial attention map, refine the input and highlight important spaces.

[0010] Preferably, the second stage introduces a conditional feature fusion module in the encoder to form a perceptual decoding residual diffusion model, which decouples the conditional information in the image and provides more structured guidance for the decoder through interpretable dynamic conditional perception; the input of the perceptual decoding residual diffusion model is represented as x in , which is the output of the multi-scale feature selection fusion network MCESS-Unet pre . input .

[0011] Preferably, the conditional feature fusion module integrates the features of x input and x pre in the perceptual encoder, and processes the features from the two inputs at the same time, and the corresponding expression is as follows:

[0012] F p = Conv 3×3 (LeakRelu(Conv 3×3 (x pre ))),

[0013] F in = Conv 3×3 (LeakRelu(Conv 3×3 (x input ))),

[0014] where Conv 3×3 represents a 3x3 convolution kernel, and LeakRelu is an activation function, and the conditional features F p and F in are obtained through such operations, and the two conditional features are connected as follows: F c = Concatenation (F p , F in ), and the connected feature is represented as F c ; then, further learn the spatially adaptive weights and biases as follows: F i = ω i (F c ) · F p + Υ i (F c ), where ω i (·) and Υ i (·) are spatially adaptive weights and biases learned from the combination of the guide feature F c and the conditional feature F p obtained previously, and i represents the feature information of the i-th layer of the decoder.

[0015] Preferably, the remote sensing image defogging method comprises the following steps:

[0016] Step 1: Establishing a defogging model;

[0017] Step 2: Establishing a two-stage defogging network framework;

[0018] Step 3: Collecting a training data set;

[0019] Step 4: Training the two-stage defogging network framework to obtain a defogging network;

[0020] Step 5: Remote sensing image defogging; inputting an image to be defogged into the network, and obtaining a restored clear image at the output end through forward propagation.

[0021] Preferably, in the first stage, the training uses a loss function L total , and the second stage uses an L1 loss, wherein L total The loss function includes the following loss functions:

[0022]

[0023] L total = L L1 + L mse + λ·L L1-scale

[0024] wherein y i is an actual output result of the network, is a corresponding true value image, L mse represents a reconstruction loss, R(·) is a 255-fold remolding of each pixel of the image to the original size, and λ is a weight of the L L1-scale corresponding loss.

[0025] The advantages of the present application are that: the present method designs two modules, one for feature extraction and one for feature selection. The MC module is used to extract features of different scales, and the ESS module then selectively weights and fuses these features, selectively retaining channel and spatial information. The two blocks complement each other, effectively extracting multi-scale features and analyzing spatial and channel information. The defogging performance is effectively enhanced. At the same time, since the present method first performs preprocessing in the first stage to obtain a rough defogging prediction image, and then performs image modification based on the prediction image, the calculation resources and time are effectively saved. The diffusion model in the current image repair research needs about two million iterations, and the improved present method only needs 300,000 iterations.

[0026] In the scheme, the multi-scale feature extraction module combines convolution with multi-scale channel attention (MSCA) mechanism to optimize feature selection and expression ability, while selectively enhancing important features and suppressing irrelevant or noise information. The output feature map of the ESS module is optimized in both channel and spatial dimensions, providing strong representation ability; and a spatial attention is set therein to compensate for the defect that some important features are ignored in the selection mechanism. The conditional feature fusion module decouples the conditional information in the image, and provides more structured guidance for the decoder through interpretable dynamic conditional perception. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A structure diagram of a two-stage dehazing network framework in a remote sensing image dehazing method based on a diffusion model.

[0028] Figure 2 A structure diagram of a multi-scale feature selection fusion network MCESS-Unet in a remote sensing image dehazing method based on a diffusion model.

[0029] Figure 3 A structure diagram of a multi-scale feature extraction module. Figure 2

[0030] A structure diagram of a selection fusion module. Figure 4 Figure 2 A structure diagram of a multi-scale feature extraction module.

[0031] Figure 5 A structure diagram of a perception decoding residual diffusion model in a remote sensing image dehazing method based on a diffusion model.

[0032] Figure 6 A basic step flowchart of the present application.

[0033] Figure 7 A visual comparison result of applying the present application and other prior art dehazing to a graph in the Haze1k-thin data set.

[0034] Figure 8 A visual comparison result of applying the present application and other prior art dehazing to a graph in the Haze1k-moderate data set.

[0035] Figure 9 A visual comparison result of applying the present application and other prior art dehazing to a graph in the Haze1k-thick data set.

[0036] Figure 10 A visual comparison result of applying the present application and other prior art dehazing to a graph in the RSID data set.

[0037] ​Figure 11 Visual comparison results of applying the present invention and other prior art dehazing for one image in the DHID dataset.

[0038] Figure 12 Visual comparison results of applying the present invention and other prior art dehazing for one image in the RainDrop dataset.

[0039] Figure 13 Visual comparison results of applying the present invention and other prior art dehazing for one image in the RainDrop dataset.

[0040] Figure 14 Visual comparison results of applying the present invention and other prior art dehazing for one image in the Realblur dataset. DETAILED DESCRIPTION

[0041] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings, and the description of the embodiments will help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solutions of the present application.

[0042] As shown in Figures 1-5 The present application provides a remote sensing image dehazing method based on a diffusion model, which adopts a two-stage dehazing network framework, and the two-stage dehazing network framework includes a multi-scale feature selection fusion network MCESS-Unet in the first stage and a perception decoding residual diffusion model in the second stage.

[0043] The first stage: the multi-scale feature selection fusion network MCESS-Unet is based on the U-net architecture and fuses the multi-scale feature extraction (MC) module and the selection fusion (ESS) module, each encoder and decoder contains a multi-scale feature extraction module, and the final output of the encoder part is output to the decoder part after being processed by the feature fusion module, while the selection fusion module is set on each jump connection.

[0044] Since the remote sensing image contains haze and multi-scale detail information, and the calculation amount of the Transformer method is large, in the first stage of feature extraction, the present method uses the attention mechanism to achieve a rough dehazing effect. The multi-scale feature extraction module is used for feature extraction, and the selection fusion module is used for feature selection. The former extracts features of different scales, and then the latter selectively weights and fuses these features to enhance the dehazing performance.

[0045] The encoder extracts hierarchical features through the multi-scale feature extraction module, while reducing the spatial dimension through max-pooling. In the decoder, the spatial resolution is gradually reconstructed using standard convolution. At each stage, the multi-scale feature extraction module captures deeper hierarchical features, while the selection fusion module is used to fuse these hierarchical features with the corresponding encoder features, replacing the traditional skip connection. The multi-scale feature extraction module enhances multi-scale features, while the selection fusion module selectively retains channel and spatial information. The two blocks complement each other, effectively extracting multi-scale features and analyzing spatial and channel information.

[0046] In addition, the method can also introduce an ASFusion module to further fuse features. The final output consists of a predicted image x pre and an auxiliary low-resolution output x0, both optimized using a hybrid loss function Loss total .

[0047] I. Multi-scale feature extraction module: In traditional convolutional neural networks, the size and stride of the convolution kernel are fixed, which means that the receptive field processed by the network at different levels is hard-coded and cannot adaptively change according to the input data. Attention mechanisms help capture these relationships, but they still struggle to handle multi-level features. The multi-scale feature extraction module combines convolution with a multi-scale channel attention (MSCA) mechanism, which can optimize feature selection and expression capabilities.

[0048] The multi-scale feature extraction module uses a multi-scale channel attention (MSCA) module, which uses different sizes of convolution kernels (i.e., 1x1 (pixel-wise), 3x3, and 5x5) and max-pooling to capture features at different scales. After tensor concatenation, a 1x1 convolution is used to integrate the output, and a channel attention (CA) mechanism is used to selectively enhance important features while suppressing irrelevant or noise information. This module improves feature extraction capabilities, as shown in Figure 3 which enhances the traditional convolutional neural network (CNN) architecture by combining multi-scale feature maps with a channel attention mechanism.

[0049] II. Selection fusion module: Simple skip connections help solve the gradient vanishing problem and speed up training, but they struggle to handle complex features and multi-scale information. To improve feature representation and fusion, the method improves the skip connection by setting a selection fusion (ESS) module that efficiently fuses features through gating and effectively retains features through a channel attention mechanism, selection mechanism, and spatial attention mechanism.

[0050] The input features of the selection fusion module include the output features of the down-sampled multi-scale feature extraction module and the output tensor of the decoder up-sampling (both have corresponding shapes), which are input 1 and input 2, respectively. These two input features are first fused with global context information through an efficient gating mechanism (EAG), and then adaptively adjusted using a channel attention (CA) mechanism to highlight key channel features. These features are then processed by a selection mechanism, which simultaneously applies channel compression and excitation (cSE) and spatial compression and excitation (sSE). Among them, cSE extracts spatial information through global average pooling, and then generates channel attention weights through a fully connected layer and a sigmoid function in sequence to recalibrate the feature map and emphasize important channels. sSE applies an activation function after a 1x1 convolution to enhance the feature map, create a spatial attention map, refine the input, and highlight important spaces. The outputs of cSE and sSE are added to the residual, and the importance maxima operator is used to select the feature value with the largest weight output. Finally, the output feature map is optimized in both channel and spatial dimensions, providing strong representation capabilities. It is worth mentioning that the selection fusion module adds spatial attention in the last selection mechanism, because the selection mechanism will suppress some important features that are ignored, so spatial attention is combined with the residual to obtain the final output features of the selection fusion module.

[0051] The input representation of the second stage is x in , the output of the MCESS-Unet pre is concatenated with the input x input . Then the image is gradually recovered by step-by-step denoising. Traditional diffusion models usually use basic up-sampling operations for decoding, which leads to a heavy reliance on the encoder output features and causes important information to be lost. This approach has limitations in recovering image details. To solve these problems and improve image recovery quality, this method improves the diffusion model by introducing a conditional feature fusion (CFF) module in the encoder, forming a new perceptual decoding residual diffusion model.

[0052] The conditional feature fusion module decouples the conditional information in the image and provides more structured guidance for the decoder through interpretable dynamic conditional perception. The feature sets x input and x pre in the image are integrated into the perceptual encoder, which simultaneously processes features from both inputs, with the corresponding expressions as follows:

[0053] F p = Conv 3×3 (LeakRelu(Conv 3×3 (x pre ))),

[0054] F in = Conv3×3 (LeakRelu(Conv 3×3 input ))),

[0055] where Conv 3×3 represents a 3x3 convolution kernel, LeakRelu is an activation function, and the conditional feature F p is obtained by such an operation in Two conditional features are concatenated: F c = Concatenation(F p , F in ), and the connected feature is denoted as F c . Then, further learn the spatial adaptive weights and biases as follows: F i = ω i (F c )·F p + Υ i (F c ), where ω i (·) and Υ i (·) are spatial adaptive weights and biases learned from the combination of the previously obtained guiding feature F c and the conditional feature F p , and i indicates that this is the feature information of the i-th layer of the decoder. The above improvement not only enhances the perceptual ability of the encoder, but also significantly improves the ability of the decoder to restore details.

[0056] Next, the scheme is verified through specific experiments:

[0057] ​The method was evaluated on eight public datasets, each with different features and purposes, including six remote sensing haze datasets, one raindrop dataset, and one blur dataset. The StateHaze1k dataset contains three subsets: thin haze, moderate haze, and heavy haze. These three subsets are combined together, containing a total of 900 pairs of images for training, and 45 pairs of images for testing in each subset. RSID is a military site dataset containing 1000 pairs of images, each with a corresponding version. 100 pairs of images were randomly selected for testing, and the remaining 900 pairs of images were used for training. The DHID dataset is designed specifically for heavy haze dehazing tasks, containing 14,490 pairs of images for training and a separate 500 pairs of test set. The LHID dataset is constructed specifically for synthetic optical dehazing, containing a training set and two independent test sets: LHID-A and LHID-B. This study uses LHID-A for testing. The RainDrop dataset was taken by two identical glasses: one glass with water sprayed on it, and the other kept clean. The dataset for image raindrop removal contains 861 training images and 58 test images. Realblur is a real blur dataset for learning and evaluating deblurring algorithms, containing 3843 pairs of training images and 1019 test images. The image size is 512x512 (except for RSID and RainDrop, where the image size is 256x256 for RSID and 720x480 for RainDrop).

[0058] In the experiment, the computer configuration was as follows: Intel(R) Xeon(R) Siler 4210R CPU processor, Nvidia RTXA5000 graphics processor, main frequency 2.40 GHz, memory 256 GB, and operating system ubuntu 20.04. The dehazing method was implemented based on the Pytorch framework. As shown in Figure 6 The present application specifically includes the following steps:

[0059] Step 1: Establish a dehazing model.

[0060] Let h represent the foggy image, r represent the restored clear image, and the function F represent the mapping relationship between the foggy image and the corresponding clear image. Then the dehazing problem is modeled as follows (i.e., the dehazing model): r = F(h). According to the above formula, once the mapping relationship F is obtained, given the foggy image h, the clear image can be obtained through the function relationship mapping, thereby realizing image dehazing.

[0061] Step 2: Establish a two-stage dehazing network framework with perceptual decoding and two-stage residual diffusion.

[0062] The two-stage defogging network framework is constructed according to the defogging model established in step 1. The input of the first stage is the foggy image, and the output is the corresponding clear image. The entire network of the first stage is an end-to-end structure. Each layer of the network extracts multi-scale features, and then the output tensor corresponding to the encoder-decoder is processed by the selection fusion to obtain the preprocessed image. In the second stage, the preprocessed image is used as the conditional input of the residual diffusion model with perceptual decoding, wherein the perceptual decoder optimizes the residual estimation by further decoupling the condition, thereby improving the generation quality.

[0063] Step 3: Collect the training data set.

[0064] The present application uses five data sets, three remote sensing haze data sets SateHaze1k, RSID and HRSD respectively. SateHaze1k includes three subsets: thin fog, medium fog and thick fog, each subset contains 300 training images and 45 test images. Thin fog images use fog masks from real clouds, while medium fog samples combine the features of fog and medium fog. Thick fog images are generated using a thick fog transmission map. RSID is a military site data set containing 1000 pairs of images, each pair has a corresponding version. We randomly select 100 pairs of images for testing, and the remaining 900 pairs of images for training. The HRSD data set is divided into two subsets: LHID and DHID. LHID contains a total of 30,517 training images and 500 test images, all of which are generated using an atmospheric scattering model. This subset simulates various haze conditions to train the network to handle different levels of haze in remote sensing images. On the other hand, DHID is composed of a total of 14,990 images, which are created using real haze maps and can more realistically present haze. Among them, 14,490 images are used for training and 500 images are used for testing. The RainDrop data set is taken by two identical glasses: one glass is sprayed with water, and the other glass remains clean. The data set for image raindrop removal contains 861 training images and 58 test images. Realblur is a real blur data set for learning and evaluating deblurring algorithms, containing 3843 pairs of training images and 1019 test images.

[0065] Step 4: Train the two-stage defogging network framework to obtain the defogging network.

[0066] Learning-based defogging methods require labeled fog samples for training. The data sets currently used more are SateHaze1k, RSID, HRSD, RainDrop and Realblur.

[0067] In this step, the network is trained with the collected foggy image samples to learn the mapping relationship between the foggy image and the clear image. In the first stage, the training uses the loss function L total , and the second stage uses the L1 loss, where L total The loss function includes the following loss functions:

[0068]

[0069] L total =L L1 +L mse +λ·L L1-scale

[0070] where y i is the actual output result of the network, is the corresponding true value image, L mse represents the reconstruction loss, R(·) is to restore each pixel of the image to the original size of 255 times, which enhances the sensitivity of the model to pixel-level differences, because pixel-level differences are easier to distinguish, and the corresponding loss is named L L1-scale , and λ is the weight of the corresponding loss L L1-scale is an adjustable parameter.

[0071] In the first stage, MCESS-Unet is trained on two NVIDIA RTX A5000 GPUs. The model is trained on the StateHaze1K, RSID, RainDrop and Realblur datasets for 300 epochs, and on the DHID and LHID datasets for 100 epochs, with a batch size of 6. We use the Adam optimizer with β1=0.9, β2=0.999, an initial learning rate of 0.0002, and decay to 0.000001 through a cosine annealing scheduler. In the second stage, we follow the experimental settings in the original RDDM paper and use a single NVIDIA RTX A5000 GPU to train RDDMPD for 300,000 iterations with a batch size of 1 and a fixed learning rate of 0.00008. We use the step-by-step jump strategy of DDIM for sampling, and obtain the final result after five steps of sampling.

[0072] Step 5: Remote sensing image dehazing.

[0073] The remote sensing image dehazing method designed by the application is end-to-end. Once the network model is trained, only the image to be dehazed needs to be input into the network, and the restored clear image can be obtained at the output end through the forward propagation of the network.

[0074] In the experiment, DS-RDMPD is compared with several advanced defogging methods on small remote sensing foggy datasets: DCP, AOD-Net, FFA-Net, Trinity-Net, Dehazeformer-b, OK-Net and AU-Net. In large remote sensing datasets, the comparison methods are: DCP, AOD-Net, FCTF-Net, GridDehaze-Net, FFA-Net, MixDehaze-Net, AU-Net and RSHazeDiff. In the rain removal task, it is also compared with several latest methods: IDT, Restormer and DRSformer. In the defogging task, it is also compared with several latest methods: Uformer, NAFNet and MLWNet. In order to ensure fair comparison, the official implementation of these deep learning models is used in the training process. PSNR represents the peak signal-to-noise ratio, SSIM represents the result similarity. Hazy Image represents the foggy image, and GT represents the real image. The performance comparison can refer to Tables 1-4. For the compared methods, the best method and the second best method are marked in bold and underlined; as follows.

[0075] Table 1: Performance table of SateHaze1k and RSID data

[0076]

[0077] Table 2: Performance table of HRSD data

[0078]

[0079] Table 3: Performance table of RainDrop data

[0080]

[0081] Table 4: Performance table of Realblur data

[0082]

[0083] The intuitive effect can be referred to Figures 7-14 .

[0084] The above describes the present application in conjunction with the drawings, and it is obvious that the specific implementation of the present application is not limited by the above manner, as long as various non-essential improvements are made by adopting the inventive concept and technical solution of the present application, or the inventive concept and technical solution of the present application is directly applied to other occasions without improvement, all of which are within the protection scope of the present application.

Claims

1. A remote sensing image defogging method based on a diffusion model, characterized by: A two-stage dehazing network framework is adopted, which includes a multi-scale feature selection fusion network MCESS-Unet in the first stage and a perceptual decoding residual diffusion model in the second stage; the multi-scale feature selection fusion network MCESS-Unet processes the input image to generate a predicted image to achieve a rough dehazing effect. The multi-scale feature selection fusion network MCESS-Unet in the first stage is based on the U-net architecture and integrates a multi-scale feature extraction module and a selective fusion module; the multi-scale feature extraction module is used to extract features of different scales, and the selective fusion module then selectively weights and fuses these features; in the second stage, the predicted image is then used as the conditional input of the perceptual decoding residual diffusion model, and after being connected with the input image, it is input into the perceptual decoding residual diffusion model. The perceptual decoding residual diffusion model optimizes the residual estimation by further decoupling the conditions, and gradually denoises and restores the image.

2. The remote sensing image defogging method based on a diffusion model according to claim 1, characterized in that: In the multi-scale feature selection fusion network MCESS-Unet, each encoder and decoder contains a multi-scale feature extraction module. The final output of the encoder part is processed by the feature fusion module and output to the decoder part, and the selection fusion module is set on each skip connection; the encoder extracts hierarchical features through the multi-scale feature extraction module and reduces the spatial dimension through maximum pooling. In the decoder, standard convolution is used to gradually reconstruct the spatial resolution. The selection fusion module is used to fuse these hierarchical features with the corresponding encoder features, thereby replacing the traditional skip connection.

3. The remote sensing image defogging method based on a diffusion model according to claim 2, characterized in that: The multi-scale feature extraction module uses convolution kernels of different sizes and maximum pooling to capture features of different scales. After tensor splicing, the output is integrated through 1×1 convolution, and the channel attention mechanism is used to selectively enhance important features.

4. The remote sensing image defogging method based on a diffusion model according to claim 1, characterized in that: The two input features of the fusion module are first fused with the global context information through an efficient gating mechanism, and then adaptively adjusted using the channel attention mechanism to highlight the key channel features; It is then processed by a selection mechanism that applies channel compression and excitation and spatial compression and excitation simultaneously; the outputs of both are added to the residual, and the eigenvalue with the largest weight is selected as the output using the importance maximum operator.

5. The remote sensing image defogging method based on a diffusion model according to claim 4, characterized in that: Channel compression and excitation extracts spatial information through global average pooling, and then generates channel attention weights through a fully connected layer and a sigmoid function in sequence; spatial compression and excitation applies an activation function after a 1×1 convolution to enhance the feature map, create a spatial attention map, refine the input, and highlight important spaces.

6. The remote sensing image defogging method based on a diffusion model according to claim 1, characterized in that: In the second stage, a conditional feature fusion module is introduced into the encoder to form a perceptual decoding residual diffusion model. The conditional feature fusion module decouples the conditional information in the image and provides more structured guidance for the decoder through interpretable dynamic conditional perception.

7. The remote sensing image defogging method based on a diffusion model according to claim 6, characterized in that: The conditional feature fusion module converts x input and x pre The features in the image are integrated into the perceptual encoder, which processes features from two inputs at the same time. The corresponding expressions are as follows: F p =Conv 3×3 (LeakRelu(Conv 3×3 (x pre ))), F in =Conv 3×3 (LeakRelu(Conv 3×3 (x input ))), Among them, Conv 3×3 Represents the convolution kernel 3×3 convolution, LeakRelu is the activation function, and the conditional feature F is obtained through such an operation p and F in , connect these two conditional features: F c =Concatenation(F p ,F in ), the connected feature is represented as F c ; Then, we further learn the weights and biases of spatial adaptation as follows: F i =ω i (F c )·F p +Υ i (F c ), where ω i (·) and γ i (·) is the guided feature F obtained from the previous c and conditional feature F p The spatial adaptive weights and biases learned from the combination of , i indicates that this is the feature information of the i-th layer of the decoder.

8. The remote sensing image defogging method based on a diffusion model according to any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: Establish a dehazing model; Step 2: Establish a two-stage dehazing network framework; Step 3: Collect training data set; Step 4: Train the two-stage dehazing network framework to obtain the dehazing network; Step 5: Remote sensing image dehazing: Input the image to be dehazed into the network, and obtain the restored clear image at the output through forward propagation.

9. The remote sensing image defogging method based on a diffusion model according to claim 8, characterized in that: In the first stage, training uses the loss function L total , the second stage uses L1 loss, where L total The loss functions include the following loss functions: L total =L L1 +L mse +λ·L L1-scale Among them, y i is the actual output of the network, is the corresponding true value image, L mse represents the reconstruction loss, R(·) is to reshape each pixel of the image to 255 times the original size, and λ is the corresponding loss L L1-scale The weight of .

Images