Method and system for defogging remote sensing image in image processing mode
By constructing a convolutional neural network model, combining SwinTransformer and loop filling method, the problems of unsatisfactory defog removal and high computational complexity in the existing technology are solved, and a high-quality and efficient remote sensing image defog removal effect is achieved.
Patent Information
- Application Number
- CN202510218330.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-05-30
AI Technical Summary
The prior art has poor fog removal effect when dealing with non-uniform haze in complex natural environments, and the calculation complexity is high during high-resolution image processing, making it difficult to meet the real-time requirements.
A method of defogging remote sensing images through image processing is proposed. By constructing a convolutional neural network model, including DeHazeBLOCK and ECABlock, combined with SwinTransformer's image processing method, the circular filling method is used to enhance the image edge features to achieve comprehensive reuse of feature extraction.
It significantly improves the quality of the fog removal effect and the real-time and robustness of the model. It is suitable for processing high-resolution images and complex scenes, and is especially suitable for tasks with high real-time requirements such as autonomous driving and drone navigation.
Smart Images
Figure CN120070241A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and particularly relates to a method and system for removing haze from remote sensing images by means of image processing. Background Art
[0002] Traditional image dehazing methods are mainly based on physical models. For example, the Dark Channel Prior (DCP) is used to estimate the atmospheric light and the atmospheric transmission map, and the clarity of the image is restored through these parameters. These methods work well when dealing with simple scenes, but in complex and variable natural environments, especially when facing non-uniform haze, their performance is often unsatisfactory. Specifically, traditional methods rely on assumptions about the haze distribution, such as the uniform haze model, while the haze distribution in actual scenes is often non-uniform, resulting in unsatisfactory dehazing effects. In addition, traditional methods have a high computational complexity when dealing with high-resolution images and are difficult to meet the real-time requirements.
[0003] With the development of deep learning technology, dehazing methods based on convolutional neural networks (CNNs) have received extensive attention due to their end-to-end training ability and powerful feature extraction ability. Compared with traditional physical models, deep learning models can automatically learn the distribution characteristics of haze from a large amount of data, thus showing stronger robustness when dealing with complex scenes. However, existing deep learning dehazing models also have some limitations. First, these models usually have a large number of parameters, resulting in a high computational cost and being difficult to deploy on resource-constrained devices. Second, although deep learning models have improved in dehazing effects, their training process requires a large amount of labeled data, and it is difficult to obtain high-quality pairs of hazy and haze-free images in practical applications. In addition, existing deep learning dehazing models still face the problem of low computational efficiency when dealing with high-resolution images and are difficult to adapt to tasks with high real-time requirements, such as scenarios like autonomous driving and drone navigation.
[0004] To solve these problems, researchers have proposed various improvement methods in recent years. For example, some studies reduce the number of model parameters by designing lightweight network structures, thereby reducing the computational cost; others try to combine the advantages of physical models and deep learning methods to improve the generalization ability of the model by introducing prior knowledge. In addition, generative adversarial networks (GANs) have also been applied to the dehazing task to generate more realistic dehazed images through adversarial training. However, these methods still face many challenges in practical applications, such as how to further improve the real-time performance and robustness of the model while ensuring the dehazing effect. Summary of the Invention
[0005] The object of the present invention is to propose a method and system for dehazing remote sensing images by means of image processing, which can improve the dehazing effect while reducing the number of model parameters and computational complexity, and restore clearer remote sensing images with higher quality.
[0006] According to the first aspect of the embodiments of the present disclosure, a method for dehazing remote sensing images by means of image processing is provided, including the following steps:
[0007] Construct a convolutional neural network model, which includes a convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are jointly input into ECABlock1;
[0008] Train the convolutional neural network model, and then use a loss function to constrain the output result, adjust the parameters of the convolutional neural network model, and obtain a remote sensing image dehazing enhancement network model;
[0009] Use the remote sensing image dehazing enhancement network model to process the foggy remote sensing image to obtain a dehazed image.
[0010] In one embodiment, the structures of DeHazeBLOCK1, DeHazeBLOCK2, DeHazeBLOCK3, DeHazeBLOCK4, and DeHazeBLOCK5 are the same.
[0011] In one embodiment, the initial features input into DeHazeBLOCK1 or DeHazeBLOCK2, DeHazeBLOCK3 or DeHazeBLOCK4 or DeHazeBLOCK5 Figure 1 pass through a normalization layer to obtain features Figure 2 The features Figure 2 are input into a cyclic padding layer to perform a cyclic padding operation on the features Figure 2 The output of the cyclic padding layer passes through a linear mapping layer 1 and a linear mapping layer 2. After the output of the linear mapping layer 2 performs a multi-head attention operation, the feature maps of the convolutional layer and the linear mapping layer 1 are concatenated channel-wise as the input of the linear mapping layer 3; after trimming the feature map output by the linear mapping layer 3, it is concatenated with the features Figure 2Perform a splicing operation to obtain features Figure 3 , the features Figure 3 Then perform a channel-wise splicing operation with features Figure 1 and input the result into a linear mapping module. The output of the linear mapping module is spliced with features Figure 3 as the final output of DeHazeBLOCK1 or DeHazeBLOCK2, DeHazeBLOCK3 or DeHazeBLOCK4 or DeHazeBLOCK5.
[0012] In one embodiment, the structures of ECABlock1 and ECABlock2 are the same.
[0013] In one embodiment, the implementation manner of ECABlock1 or ECABlock2 is as follows: First, perform global average pooling operation on the input features Figure 1 , features Figure 2 , features Figure 3 to compress the spatial dimension. Secondly, perform one-dimensional convolution operation on the channel feature vector after global average pooling. Next, use the Sigmoid function to normalize the feature vector to generate a channel attention weight vector. Finally, multiply the channel attention weight vector with features Figure 1 , features Figure 2 , features Figure 3 channel by channel to obtain a weighted feature map.
[0014] In one embodiment, the training process of the convolutional neural network model is as follows:
[0015] Step a: Input the hazy remote sensing image x into a convolutional layer 1 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O1 = F1(x);
[0016] Step b: Input the output feature O1 into DehazeBLOCK1: O2 = DehazeBLOCK1(O1);
[0017] Step c: Input the output feature O2 into a convolutional layer 2 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O3 = F2(O2);
[0018] Step d: Input the output feature O3 into DehazeBLOCK2: O4 = DehazeBLOCK2(O3);
[0019] Step e: Input the output feature O4 into a convolutional layer 3 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O5 = F3(O4);
[0020] Step f: Input the output feature O5 into DehazeBLOCK3: O6 = DehazeBLOCK3(O5);
[0021] Step g: Upsample and fill the output feature O6 using the nearest neighbor interpolation method: O7 = UpSample1(O6);
[0022] Step h: Input the output features O7, O2, and O4 into ECABlock1 for feature fusion: O8 = ECA(O2, O4, O7);
[0023] Step i: Input the output feature O8 into DehazeBLOCK4: O9 = DehazeBLOCK4(O8);
[0024] Step j: Upsample and fill the output feature O9 using the nearest neighbor interpolation method: O10 = UpSample2(O9);
[0025] Step k: Input the output features O10, O2, and O4 into the ECABlock2 module for feature fusion: O11 = ECA(O2, O4, O10);
[0026] Step l: Input the output feature O11 into DehazeBLOCK5: O12 = DehazeBLOCK5(O11);
[0027] Step m: Input the output feature O12 into a convolutional layer 4 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O13 = F4(O12):
[0028] Step n: Input the output feature O13 into the soft reconstruction layer to obtain the final dehazed image.
[0029] In one embodiment, the loss function is:
[0030]
[0031] where LOSS L1 represents the sum of the absolute values of minimizing the difference L1 between the true value yi and the predicted value f(xi).
[0032] According to the second aspect of the embodiments of the present disclosure, a remote sensing image dehazing system in an image processing manner is provided, including:
[0033] Convolutional neural network model, the convolutional neural network model includes a convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are jointly input into ECABlock1;
[0034] Training module, training the convolutional neural network model, and then using a loss function to constrain the output result, adjusting the parameters of the convolutional neural network model to obtain a remote sensing image dehazing enhancement network model;
[0035] Dehazing module, using the remote sensing image dehazing enhancement network model to process the foggy remote sensing image to obtain a dehazed image.
[0036] According to the third aspect of the embodiments of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program running on the memory. When the processor executes the program, it implements the method for dehazing remote sensing images in an image processing manner as described above.
[0037] According to the fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, it implements the method for dehazing remote sensing images in an image processing manner as described above.
[0038] The above technical solutions adopted by the present invention, compared with the prior art, have the following advantages: (1) Applicable to the situation where the model fully learns the characteristics of remote sensing foggy images. By proposing a U-Net-like network structure, combining the image processing method of SwinTransformer, and introducing a special feature transfer mechanism, the present invention can effectively solve the problems of significant reduction in the feature space and loss of important details and features during the feature extraction process by the model. This improved strategy realizes the full reuse of features inside the network, that is, each layer can receive the feature information of all previous layers, thereby enhancing the propagation ability of features in the network. This design is particularly applicable to the situation where the model fully learns the characteristics of remote sensing foggy images and can significantly improve the integrity and accuracy of feature extraction.
[0039] (2) Applicable to the situation of defogging foggy images with uneven fog distribution. The present invention uses the cyclic padding method to pad the foggy images, enhancing the participation of image edge or corner features in visual tasks. Specifically, the cyclic padding method copies the edge part in the opposite direction of the side to be padded to the outside of that side, ensuring that when calculating attention, the image pixels of the two parts come from different partitions. This method promotes the interaction between the features of different image partitions, thereby improving the accuracy of feature extraction and the overall performance of the model, and is particularly suitable for processing foggy images with uneven fog distribution.
[0040] (3) Applicable to the situation of expanding the clear image dataset
[0041] By processing the defogged images as clear images, the present invention enables the deep learning model to efficiently process foggy image data visually, avoiding the inefficient process of defogging with traditional physical models. This method can generate clear data for training in various fields that require remote sensing datasets (such as object detection, scene classification, etc.), providing dataset expansion support for various task models, thereby significantly improving the training effect and generalization ability of the models. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] The specification drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application.
[0043] Figure 1 It is a flowchart of the method for defogging remote sensing images in an image processing manner;
[0044] Figure 2 It is a network architecture diagram of the method for defogging remote sensing images in an image processing manner;
[0045] Figure 3 It is a DeHazeBLOC diagram in the method for defogging remote sensing images in an image processing manner;
[0046] Figure 4 It is an ECABlock diagram in the method for defogging remote sensing images in an image processing manner.
[0047] Figure 5 It is a schematic diagram of the cyclic padding method in the method for defogging remote sensing images in an image processing manner. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0048] The present disclosure will be further described below in conjunction with the drawings and embodiments.
[0049] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this application belongs.
[0050] Note that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0051] Note that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of methods and systems according to various embodiments of the present disclosure. It should be noted that each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, program segment, or part of code may include one or more executable instructions for implementing the logical functions specified in the respective embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. Similarly, it should be noted that each block in the flowchart and / or block diagram, and the combinations of blocks in the flowchart and / or block diagram, may be implemented using a dedicated hardware-based system for performing the specified functions or operations, or may be implemented using a combination of dedicated hardware and computer instructions.
[0052] Embodiment 1:
[0053] As Figure 1 shown, this embodiment provides a method for removing haze from remote sensing images by means of image processing, including the following steps:
[0054] Step 1, construct a convolutional neural network model, the convolutional neural network model includes a convolutional layer, a dehazing module DeHazeBLOCK, an efficient channel attention module ECABlock, an upsampling layer, and a soft reconstruction layer, as Figure 2As shown in the figure. Specifically, it includes a convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and a soft reconstruction layer arranged in sequence. The outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are jointly input into ECABlock1.
[0055] As Figure 3 shown, the structures of DeHazeBLOCK1, DeHazeBLOCK2, DeHazeBLOCK3, DeHazeBLOCK4, and DeHazeBLOCK5 are the same. In each DeHazeBLOCK, the initial features Figure 1 pass through a normalization layer to obtain features Figure 2 , and the features Figure 2 are input into a cyclic padding layer. The output of the cyclic padding layer passes through a linear mapping layer 1 and a linear mapping layer 2. After the output of the linear mapping layer 2 undergoes a multi-head attention operation, it is concatenated with the feature map of the convolutional layer and the linear mapping layer 1 in a channel-wise manner as the input of the linear mapping layer 3; after trimming the feature map output by the linear mapping layer 3, it is concatenated with the features Figure 2 to obtain features Figure 3 . The features Figure 3 are then concatenated with the features Figure 1 in a channel-wise manner and input into a linear mapping module. The output of the linear mapping module is concatenated with the features Figure 3 as the final output of DeHazeBLOCK.
[0056] As Figure 4 shown, the structures of ECABlock1 and ECABlock2 are the same. In each ECABlock, first, the input features Figure 1 , features Figure 2 , and features Figure 3 are compressed in the spatial dimension through a global average pooling operation. Secondly, the channel feature vector after global average pooling is processed through a one-dimensional convolutional operation. Next, the Sigmoid function is used to normalize the feature vector to generate a channel attention weight vector. Finally, the channel attention weight vector is multiplied with the features Figure 1 , features Figure 2 , and features Figure 3 channel by channel to obtain a weighted feature map.
[0057] Step 2: Train the convolutional neural network model, and then use a loss function to constrain the output results and adjust the parameters of the convolutional neural network model to obtain a remote sensing image dehazing and enhancement network model;
[0058] Specifically, in this embodiment, the dataset is the remote sensing dataset StateHaze1K, which consists of 2,400 JPG images with a resolution of 2m / pixel and a size of 400×400. The dataset is input into the convolutional neural network model for 30 times of training in the following process:
[0059] Step a: Input the hazy remote sensing image x into a convolutional layer 1 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O1 = F1(x);
[0060] Step b: Input the output feature O1 into DehazeBLOCK1: O2 = DehazeBLOCK1(O1);
[0061] Step c: Input the output feature O2 into a convolutional layer 2 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O3 = F2(O2);
[0062] Step d: Input the output feature O3 into DehazeBLOCK2: O4 = DehazeBLOCK2(O3);
[0063] Step e: Input the output feature O4 into a convolutional layer 3 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O5 = F3(O4);
[0064] Step f: Input the output feature O5 into DehazeBLOCK3: O6 = DehazeBLOCK3(O5);
[0065] Step g: Upsample and fill the output feature O6 using the nearest neighbor interpolation method: O7 = UpSample 1(O6);
[0066] Step h: Input the output feature O7 and O2, O4 into ECABlock1 for feature fusion: O8 = ECA(O2, O4, O7);
[0067] Step i: Input the output feature O8 into DehazeBLOCK4: O9 = DehazeBLOCK4(O8);
[0068] Step j: Upsample and fill the output feature O9 using the nearest neighbor interpolation method: O10 = UpSample 2(O9);
[0069] Step k: Input the output features O10, O2, and O4 into the ECABlock2 module for feature fusion: O11 = ECA(O2, O4, O10);
[0070] Step l: Input the output feature O11 into DehazeBLOCK5: O12 = DehazeBLOCK5(O11);
[0071] Step m: Input the output feature O12 into a convolutional layer 4 with a convolutional kernel size of 3x3, a stride of 2, and a padding size of 2: O13 = F4(O12):
[0072] Step n: Input the output feature O13 into the soft reconstruction layer to obtain the final dehazed image.
[0073] Step 3: Use the remote sensing image dehazing and enhancement network model to process the hazy remote sensing image to obtain a dehazed image.
[0074] In the above way, by combining the two fields of dehazing and super-resolution, the network can enhance the dehazed remote sensing image while dehazing, obtaining high-quality remote sensing images to provide high-quality data for subsequent remote sensing image processing work.
[0075] Embodiment 2:
[0076] This embodiment provides a remote sensing image dehazing system in an image processing manner, including:
[0077] A convolutional neural network model, which includes a convolutional layer 1, DehazeBLOCK1, convolutional layer 2, DehazeBLOCK2, convolutional layer 3, DehazeBLOCK3, UpSample1, ECABlock1, DehazeBLOCK4, UpSample2, ECABlock2, DehazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DehazeBLOCK1, DehazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DehazeBLOCK1, DehazeBLOCK2, and UpSample1 are jointly input into ECABlock1;
[0078] A training module that trains the convolutional neural network model, then uses a loss function to constrain the output results, and adjusts the parameters of the convolutional neural network model to obtain a remote sensing image dehazing and enhancement network model;
[0079] A dehazing module that uses the remote sensing image dehazing and enhancement network model to process the hazy remote sensing image to obtain a dehazed image.
[0080] When the prior art performs image dehazing, due to the lack of information interaction between shallow features and deep features, the feature information extraction is insufficient, and the use of circular padding leads to the loss of feature information at the edges of the feature map; the present invention proposes a U-Net-like neural network structure to perform dehazing on remote sensing images. For the image padding method in the dehazing main module DehazeBlock, the circular shift reflection padding method is used to pad the image, and then it is input into the network for dehazing operation, which can achieve an efficient dehazing effect.
[0081] Example 3:
[0082] An electronic device includes a memory, a processor, and a computer program running on the memory. When the processor executes the program, it implements the above-mentioned method for dehazing remote sensing images by image processing, including:
[0083] Construct a convolutional neural network model, which includes convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are jointly input into ECABlock1;
[0084] Train the convolutional neural network model, and then use a loss function to constrain the output result, adjust the parameters of the convolutional neural network model, and obtain a remote sensing image dehazing enhancement network model;
[0085] Use the remote sensing image dehazing enhancement network model to process the foggy remote sensing image to obtain a dehazed image.
[0086] Example 4:
[0087] A computer-readable storage medium stores a computer program, and when the program is executed by a processor, it implements the above-mentioned method for dehazing remote sensing images by image processing, including:
[0088] Build a convolutional neural network model, where the convolutional neural network model includes a convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are jointly input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are jointly input into ECABlock1;
[0089] Train the convolutional neural network model, and then use a loss function to constrain the output result, adjust the parameters of the convolutional neural network model, and obtain a remote sensing image dehazing and enhancement network model;
[0090] Use the remote sensing image dehazing and enhancement network model to process a foggy remote sensing image to obtain a dehazed image.
[0091] Those skilled in the art should understand that the above-mentioned modules or steps of the present disclosure can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple of them can be fabricated into a single integrated circuit module. The present disclosure is not limited to any specific combination of hardware and software.
[0092] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
[0093] Although the specific implementation manners of the present disclosure are described above in conjunction with the drawings, it is not a limitation to the protection scope of the present disclosure. Those skilled in the art should understand that based on the technical solutions of the present disclosure, various modifications or deformations that do not require creative labor by those skilled in the art are still within the protection scope of the present disclosure.
Claims
1. A method for defogging a remote sensing image by image processing, characterized in that: The following steps are involved: Constructing a convolutional neural network model, the convolutional neural network model includes convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2 and UpSample2 are input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2 and UpSample1 are input into ECABlock1; The convolutional neural network model is trained, and then the output result is constrained by using a loss function, and the convolutional neural network model parameters are adjusted to obtain a remote sensing image defogging enhancement network model; The remote sensing image defogging enhancement network model is used to process the foggy remote sensing image to obtain a defogging image.
2. According to claim 1, a method for defogging a remote sensing image by image processing, characterized in that: The structures of DeHazeBLOCK1, DeHazeBLOCK2, DeHazeBLOCK3, DeHazeBLOCK4 and DeHazeBLOCK5 are the same.
3. A remote sensing image defogging method using image processing according to claim 1 or 2, characterized in that: The initial feature map 1 input to DeHazeBLOCK1 or DeHazeBLOCK2, DeHazeBLOCK3 or DeHazeBLOCK4 or DeHazeBLOCK5 is passed through the normalization layer to obtain the feature map 2, the feature map 2 is input to the recurrent padding layer, the output of the recurrent padding layer passes through the linear mapping layer 1 and the linear mapping layer 2, the output of the linear mapping layer 2 is subjected to a multi-head attention operation, and then passes through the convolution layer and the feature map of the linear mapping layer 1 for channel-by-channel splicing operation as the input of the linear mapping layer 3; the feature map output by the linear mapping layer 3 is trimmed and then spliced with the feature map 2 to obtain the feature map 3, the feature map 3 is then spliced with the feature map 1 by channel and input to the linear mapping module, the output of the linear mapping module is spliced with the feature map 3 as the final output of DeHazeBLOCK1 or DeHazeBLOCK2, DeHazeBLOCK3 or DeHazeBLOCK4 or DeHazeBLOCK5.
4. The method for defogging a remote sensing image by image processing according to claim 1, characterized in that: The ECABlock1 and ECABlock2 have the same structure.
5. A remote sensing image defogging method using image processing according to claim 1 or 4, characterized in that: The implementation method of ECABlock1 or ECABlock2 is: first, the spatial dimension of the input feature map 1, feature map 2, and feature map 3 is compressed through a global average pooling operation, and then the channel feature vector after global average pooling is processed through a one-dimensional convolution operation. Next, the feature vector is normalized using the Sigmoid function to generate a channel attention weight vector, and finally the channel attention weight vector is multiplied by the feature map 1, feature map 2, and feature map 3 channel by channel to obtain a weighted feature map.
6. The method for defogging a remote sensing image by image processing according to claim 1, characterized in that: The training process of the convolutional neural network model is: Step a: Input the foggy remote sensing image x into a convolutional layer 1 with a convolution kernel size of 3x3, a step size of 2, and a padding size of 2: O1=F1(x); Step b: Input the output feature O1 into DeHazeBLOCK1: O2 = DehazeBLOCK1 (O1); Step c: Input the output feature O2 into the convolution layer 2 with a convolution kernel size of 3x3, a step size of 2, and a padding size of 2: O3 = F2(O2); Step d: Input the output feature O3 into DeHazeBLOCK2: O4=DehazeBLOCK2(O3); Step e: Input the output feature O4 to the convolution layer 3 with a convolution kernel size of 3x3, a step size of 2, and a padding size of 2: O5 = F3 (O4); Step f, input the output feature O5 into DeHazeBLOCK3: O6 = DehazeBLOCK3 (O5); Step g: upsample and fill the output feature O6 using the nearest neighbor interpolation method: O7 = UpSample 1 (O6); Step h: Input the output feature O7, O2 and O4 into ECABlock1 for feature fusion: O8 = ECA (O2, O4, O7); Step i: Input the output feature O8 into DeHazeBLOCK4: O9 = DehazeBLOCK4 (O8); Step j: upsample and fill the output feature O9 using the nearest neighbor interpolation method: O10 = UpSample 2 (O9); Step k: Input the output features O10, O2 and O4 into the ECABlock2 module for feature fusion: O11 = ECA (O2, O4, O10); Step 1: Input the output feature O11 into DeHazeBLOCK5: O12 = DehazeBLOCK5 (O11); Step m: Input the output feature O12 into a convolution layer 4 with a convolution kernel size of 3x3, a step size of 2, and a padding size of 2: O13 = F4 (O12): Step n: input the output feature O13 into the soft reconstruction layer to obtain the final dehazed image.
7. The method for defogging a remote sensing image by image processing according to claim 1, characterized in that: The loss function is: Among them, LOSS L1 It means minimizing the sum of the absolute values of the difference L1 between the true value yi and the predicted value f(xi).
8. A remote sensing image defogging system using image processing, characterized in that: include: A convolutional neural network model, comprising a convolutional layer 1, DeHazeBLOCK1, convolutional layer 2, DeHazeBLOCK2, convolutional layer 3, DeHazeBLOCK3, UpSample1, ECABlock1, DeHazeBLOCK4, UpSample2, ECABlock2, DeHazeBLOCK5, convolutional layer 4, and a soft reconstruction layer arranged in sequence; the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample2 are input into ECABlock2, and the outputs of DeHazeBLOCK1, DeHazeBLOCK2, and UpSample1 are input into ECABlock1; A training module is used to train the convolutional neural network model, and then a loss function is used to constrain the output result, and the parameters of the convolutional neural network model are adjusted to obtain a remote sensing image defogging enhancement network model; The defogging module uses the remote sensing image defogging enhancement network model to process the foggy remote sensing image to obtain a defogging image.
9. An electronic device comprising a memory, a processor and a computer program stored and running on the memory, characterized in that: When the processor executes the program, the method for defogging a remote sensing image by image processing is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a method for defogging a remote sensing image by image processing is implemented.