Image defogging system and method for low-altitude scene
Through the U-shaped network architecture and multi-level information processing, the problems of non-uniform fog distribution and distortion of defogging results in low-altitude image defogging are solved, and a more efficient image restoration effect is achieved.
Patent Information
- Application Number
- CN202511086079.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-09-05
- Estimated Expiration
- 2045-08-05
AI Technical Summary
Image dehazing technology in low-altitude scenes faces problems such as non-uniform fog distribution, insufficient utilization of frequency domain information, distortion of dehazing results under specific conditions, insufficient data and generalization. Existing methods are difficult to effectively restore clear images.
A U-shaped network architecture is adopted, combining the frequency domain fusion module, the space and channel interaction module and the physical perception module. Through wavelet transform processing and atmospheric scattering model, the frequency domain and spatial information are processed in different hierarchical modules respectively, and the attention mechanism and physical parameters are used for image restoration.
It improves the local clarity and global consistency of images in low-altitude scenes, solves the problem of distortion of defogging results in non-uniform fog distribution and large areas of haze, and improves the defogging performance of the model.
Smart Images

Figure CN120598820A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image restoration based on computer vision, and in particular relates to an image defogging system and method for low-altitude scenes. Background Art
[0002] With the development of low-altitude technology, low-altitude operations have been widely used in fields such as logistics and transportation, agricultural monitoring, and environmental protection. However, during low-altitude missions, especially under complex fog and haze conditions, image quality can be severely affected by atmospheric particulate matter, resulting in reduced image contrast, blurred details, and even complete loss of visibility of target objects. These issues directly impact the accuracy and reliability of key tasks such as navigation, target detection, and object recognition. Therefore, effectively removing the interference of fog and haze on low-altitude images and restoring clear, fog-free images has become an important research direction in the field of low-altitude visual perception.
[0003] Image dehazing technology, as a core approach to addressing this problem, has garnered widespread attention from scholars both domestically and internationally in recent years. In particular, dehazing techniques combined with deep learning, through their modeling of image degradation and data-driven feature learning, have demonstrated significant potential for low-altitude applications. However, due to the unique characteristics of low-altitude imaging, such as diverse viewpoints, complex environments, low resolution, noise interference, and uneven fog distribution, image dehazing in low-altitude scenes still faces numerous challenges.
[0004] Traditional image dehazing methods rely on atmospheric scattering models, estimating the atmospheric light value A and the transmission map t(x) to restore a haze-free image based on physical principles. These methods have some application value in low-altitude scenes, but they have significant limitations in complex scenes. Early research on image dehazing was led by the Dark Channel Prior (DCP) method proposed by He et al. in 2009. DCP assumes that at least one pixel value close to zero exists in a local region of a haze-free image. Using this prior, the transmission map is estimated and the haze-free image is restored. This method is computationally simple and effective, but performs poorly in smooth areas such as the sky. Subsequently, Fattal proposed a dehazing method based on color consistency, which enhances the accuracy of dehazing by analyzing the independence of scene radiance and transmittance.
[0005] To improve dehazing performance, many researchers have refined parameter estimation methods for atmospheric scattering models. For example, methods based on color linearization priors and minimum and maximum value priors improve model robustness by introducing new assumptions. Furthermore, optimization-based methods (such as Laplace regularization and total variation optimization) are widely used to smooth transmission maps, resulting in more natural-looking edges and details in dehazed images.
[0006] While current methods for dehazing low-altitude images have improved the quality of reconstructed images, they still face several key challenges. The first is non-uniform fog distribution: When low-altitude operations are conducted under complex terrain and weather conditions, fog often exhibits a non-uniform distribution, rendering the prior assumptions of traditional physics-based models ineffective and making it difficult for deep learning models to accurately model the problem.
[0007] Secondly, frequency domain information is not fully utilized: In the process of low-altitude image dehazing, the effective use of frequency domain information remains a challenge. Many existing methods rely primarily on spatial domain features when processing images, ignoring the important information contained in the frequency domain.
[0008] Then there's the issue of dehazing results being distorted under certain conditions: Dehazing algorithms often experience distortion when processing images containing large areas of sky or sea. The relatively uniform color and brightness distribution in these areas makes it difficult for the algorithm to accurately estimate transmittance and the effects of fog, resulting in color deviation or loss of detail.
[0009] Finally, there is the problem of insufficient data and generalization: there are currently few real dehazing datasets for low-altitude scenes, and existing models mostly rely on synthetic data for training. However, the distribution difference between synthetic data and real data will lead to a decline in the performance of the model in actual applications. Summary of the Invention
[0010] To address the above problems, the present invention provides, in its first aspect, an image defogging system for low-altitude scenes. The system is constructed based on a U-shaped network. By utilizing the different sensitivities of the U-shaped network to frequency domain information, different frequency components are directed to different layers, and a frequency domain fusion module, a space and channel interaction module, and a physical perception module are designed in the shallow, middle, and deep layers, respectively. The frequency domain fusion module takes as input the three high-frequency components of the feature map processed by wavelet transform. After being processed by the global attention mechanism of the frequency domain features, it outputs a global frequency domain similarity map, which is used to capture the similarity between pixels in the foggy degraded image and use the clearer light fog areas in the image to assist in restoring the image in the dense fog area. The input of the spatial and channel interaction module is the feature map after frequency domain fusion. The spatial attention and channel attention of the input feature map are first calculated and connected in the channel dimension. Then, random shuffling and group convolution operations are performed to reduce the uneven distribution of haze images at the channel level. The physical perception module is designed based on the atmospheric scattering model. The input is the low-frequency component of the feature map processed by wavelet transform. The low-frequency component information is used as the key parameters of the model. The dehazed and restored image is reconstructed using the atmospheric diffusion model formula in the feature space simulation.
[0011] Preferably, the frequency domain fusion module partially uses an improved self-attention module to perform feature extraction and strengthen attention to key frequency domain features; The spatial and channel interaction module fuses spatial and channel information from global average pooling and global maximum pooling on a multi-scale structure. The channel information undergoes a random shuffling operation to facilitate cross-channel frequency domain information exchange, while the spatial level is responsible for highlighting areas with significant haze distribution. The physical perception module learns physical parameters through the network, including global atmospheric light and medium transmittance , and combines low-frequency component information to estimate physical parameters and provide physical constraints for the dehazing task.
[0012] Preferably, the frequency domain fusion module is specifically: The encoder part of the network decomposes the input features in the frequency domain through discrete wavelet transform to obtain four frequency domain components. Three high-frequency components, namely the horizontal high-frequency sub-band LH, the vertical high-frequency sub-band HL and the diagonal high-frequency sub-band HH, are selected as the input of the module. The frequency domain fusion module contains three branches. The first branch connects LH, HL and HH, and then performs a connection operation. The second branch uses the horizontal high-frequency texture information LH and the vertical high-frequency texture information for attention calculation. First, the feature dimension is reorganized and the 1×1 convolution is performed to obtain the channel adjustment feature map of the horizontal high-frequency sub-band LH, the vertical high-frequency sub-band HL and the diagonal high-frequency sub-band HH. 、 and , then pass through Normalization obtains an attention matrix of size C×C, where C is the number of channels of the input feature. The third branch uses the HH component to multiply the attention matrix obtained by the second branch to obtain the weighted feature result, and then performs an element-wise addition operation with the residual obtained by the first branch to form a frequency domain fusion feature representation.
[0013] Preferably, the space and channel interaction module is specifically: It contains two branches that learn spatial information and channel information respectively. The two parts are merged in the channel dimension and then randomly shuffled. For the spatial information branch, the feature map output by the previous layer is used to extract the global features of the spatial dimension using global average pooling (GAP) and global maximum pooling (GMP), and then a concatenation operation is performed. Then, a 3×3 convolution operation is used to learn the spatial attention weight to highlight the areas with significant haze distribution. For the channel information branch, GAP and GMP are used to extract the global features of the channel dimension and then a concatenation operation is performed. Then, the weights between channels are learned through 1×1 convolution to enhance the modeling of different semantic features. The spatial information branch and the channel information branch are added at the element level, and finally a channel shuffle operation is performed to realize the feature interaction between spatial attention and channel attention. The input feature map is , where the number of channels, height and length of the feature map are C, H and W respectively. The generation process of channel attention first performs an operation on the input feature map Perform global average pooling to generate global features in the spatial dimension , using 1×1 convolution Perform feature compression and pass Activation function generates channel attention weight ; First, input feature map Global average pooling and global maximum pooling operations are performed on the channel dimension, the two features are multiplied element by element, and then the spatial attention weights are generated through 3×3 convolution.
[0014] Preferably, the space and channel interaction module also introduces a channel shuffling operation to redistribute information between channels, divide the channels into multiple groups, perform shuffling operations within each group of channels, and achieve rearrangement and interaction of cross-channel information through the channel shuffling operation; the weights after channel shuffling are processed by 3×3 convolution to generate the final channel-specific spatial importance map, and the importance map is normalized using the sigmoid activation function to obtain the final attention map.
[0015] Preferably, the formula of the atmospheric diffusion model in the physical perception module is:
[0016] In the formula Represents a haze image, represents a clean image without fog, represents the medium transmittance in the atmosphere, Represents the atmospheric light parameters, and the process of reconstructing the defogging recovery result is to use model learning and These two key physical parameters ultimately result in a more accurate fog-free image ; The input of the module is the deepest layer of the network, and the input is the feature map of the previous level , which contains two branches to reversely deduce and estimate the global atmospheric light and medium transmittance ; For estimating atmospheric light The branch first passes the global average pooling Learn the overall brightness of the image, and then obtain a feature map with more useful features through dimensionality increase and dimensionality decrease convolution operations, and then use Get the final estimated value; finally, use the estimated physical parameters and the atmospheric scattering model formula to calculate the approximate clear image.
[0017] A second aspect of the present invention provides an image defogging method for low-altitude scenes, comprising the following steps: Shoot and acquire low-altitude scene images; Inputting the image into the low-altitude scene image defogging system as described in the first aspect to perform object segmentation; Output the target image after dehazing.
[0018] Compared with the prior art, the present invention has the following beneficial effects: The system structure designed by the present invention is improved based on a U-shaped architecture including three layers of upsampling and downsampling, and mainly includes three parts: an attention-based frequency domain fusion module, a space and channel interaction module, and a physical perception module. The shallow attention-based frequency domain fusion module performs frequency domain decomposition and fusion of input features through discrete wavelet transform in the encoder part of the network. The fused frequency domain features contain both global low-frequency background information and local high-frequency details, providing a richer feature representation for subsequent modules. The image of the space and channel interaction module in the middle layer captures important features of the spatial dimension and the channel dimension at the same time, significantly improving the local clarity and global consistency of the dehazed image. The deep physical perception module combined with the prior information of the physical model can provide stronger constraints and interpretability for the dehazing task. The present invention solves the problem of color distortion of dehazing results in non-uniform fog distribution and large-scale haze areas in low-altitude scenes, and improves the modeling ability of the model for the degradation characteristics of haze images. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate the technical solutions of the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, what is described below is only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0020] Figure 1 This is the overall system structure diagram of the present invention.
[0021] Figure 2 It is a structural diagram of the frequency domain fusion module of the present invention.
[0022] Figure 3 This is a schematic diagram of the space and channel interaction module structure of the present invention.
[0023] Figure 4 This is a structural diagram of the physical perception module of the present invention.
[0024] Figure 5 Result diagram of the defogging effect of the model of the present invention in the embodiment. DETAILED DESCRIPTION
[0025] The invention will be further described below with reference to specific embodiments.
[0026] The data of the haze images used in this embodiment come from drone photography and the Reside public dataset, where the drone photography dataset uses the DJI M3T drone as the image acquisition device, which is equipped with a full-frame camera and a high-sensitivity sensor with an effective pixel of 45 million. In this embodiment, a drone is set up to fly and shoot according to a specific route, which consists of 51 waypoints. A total of 4239 images were obtained when the drone flew at altitudes of 300 meters, 600 meters, and 900 meters. The resolution of each image in the dataset is 4000×3000 pixels. There are slight errors in the pictures taken by the drone at the same waypoint. This is because the GPS positioning of the drone cannot be particularly accurate, so it is impossible to capture paired foggy images and fog-free image data for training. To address this problem, this embodiment only captures fog-free images as the baseline truth value, and obtains haze images through depth information synthesis, which is specifically implemented as follows: The MonoDepth2 deep learning model was selected and pre-trained on public depth datasets (such as KITTI and NYU DepthV2). It was then fine-tuned using transfer learning for fog-free image scenes taken by drones.
[0027] Depth map generation process: The fog-free image is input into the trained depth estimation model. After being processed by the model's internal feature extraction layer and depth prediction layer, combined with post-processing optimization (using a bilateral filtering algorithm to smooth and optimize the prediction results), the depth map corresponding to the input image pixel is output. Each pixel value in the depth map represents the actual distance between the corresponding scene and the drone camera. For each pixel of the fog-free image J(x), the atmospheric scattering model formula is substituted to calculate:
[0028] That is, according to the depth value of the pixel Calculate transmittance , combined with the set atmospheric light value =0.6, solve to get the pixel value corresponding to the haze image By performing pixel-by-pixel operations, we generated a complete simulated haze image, achieving simulations of haze effects with different concentrations and distribution characteristics. Based on the 4239 clean, haze-free images we captured, we obtained the corresponding haze images using the above method, resulting in a total of paired low-altitude haze training datasets.
[0029] Aiming at the problem of image dehazing modeling for real scenes, the impact of haze in low-altitude scene images is manifested in the frequency domain as the loss of high-frequency details and the blurring of low frequencies. Therefore, the present invention designs a multi-level modeling mechanism based on frequency domain characteristics, introduces different frequency information into different level modules of the U-shaped network, and enhances the network's modeling ability for details and global information. It mainly includes an attention-based frequency domain fusion module, a space and channel interaction module, and a physical perception module. The overall structure is as follows: Figure 1 As shown in Figure 2, the hierarchy has three levels, in which different blocks are adopted to extract corresponding features.
[0030] The shallow structure adopts the frequency domain fusion module based on attention, such as Figure 2 As shown, the encoder portion of the network uses a discrete wavelet transform to decompose the input features in the frequency domain, using four kernels to decompose the input features into four subbands. LL: The low-frequency subband contains the image's background information and global structural features. LH, HL, and HH: The high-frequency subbands represent high-frequency texture information in the horizontal, vertical, and diagonal directions, respectively. For the high-frequency subbands (LH, HL, and HH), this paper designs an attention-based fusion mechanism, using one-dimensional convolution and depthwise convolution to learn the local dependencies between high-frequency features. The attention mechanism generates a weight map for high-frequency features to guide the selection of important high-frequency information. Low-frequency feature enhancement: A 1×1 convolution compresses low-frequency information to extract global features. The low-frequency features are combined with the high-frequency features to form a fused frequency-domain feature representation. First, the three frequency-domain features are element-wise multiplied to obtain a new feature representation. Then, a 1×1 convolution is applied to this new feature representation to further extract and fuse features. Next, a dot product operation is performed on the feature representations of LH and HL, and an activation function is applied to generate attention weights. These weights are used to adjust the feature representation of HH. The adjusted HH feature representation is added to the feature representation obtained previously by 1×1 convolution to obtain the final fusion feature.
[0031] The fused frequency domain features contain both global low-frequency background information and local high-frequency details, providing richer feature representation for subsequent modules. The specific process is expressed by the following formula, where Represents the channel dimension transformed by 1×1 convolution, Represents a connection operation in the channel dimension. 、 and Represents the adjusted feature maps of the horizontal high-frequency sub-band LH, the vertical high-frequency sub-band HL, and the diagonal high-frequency sub-band HH:
[0032] The fused frequency domain features contain global low-frequency background information while retaining local high-frequency details, providing richer feature representation for subsequent modules.
[0033] The middle layer designs multi-scale spatial channel information fusion using space and channel interaction modules, such as Figure 3 As shown in the figure, the module can capture important features of spatial and channel dimensions at the same time, especially in scenes with non-uniform haze distribution, significantly improving the local clarity and global consistency of the dehazed image. For spatial information, global average pooling (GAP) and global maximum pooling (GMP) are used to extract global features of the spatial dimension, and spatial attention weights are learned through 3×3 convolution operations to highlight areas with significant haze distribution. For channel information, GAP and GMP are used to extract global features of the channel dimension, and the weights between channels are learned through 1×1 convolution to enhance the modeling of different semantic features. Finally, the channel shuffle operation is used to realize the feature interaction between spatial attention and channel attention. The input feature map is , where the number of channels, height and length of the feature map are C, H and W respectively. The generation process of channel attention first performs an operation on the input feature map Perform global average pooling to generate global features in the spatial dimension , using 1×1 convolution Perform feature compression and pass Activation function generates channel attention weight ; First, input feature map Global average pooling and global maximum pooling operations are performed on the channel dimension, the two features are multiplied element by element, and then the spatial attention weights are generated through 3×3 convolution.
[0034] To further enhance information interaction between spaces and channels, this module introduces a channel shuffling operation to redistribute information across channels, preventing the attention mechanism from focusing solely on local features and lacking global information integration. Channels are divided into multiple groups, and a shuffling operation is performed within each group. This channel shuffling operation enables the rearrangement and interaction of cross-channel information. The shuffled weights are processed through a 3×3 convolution to generate the final channel-specific spatial importance map. This importance map is then normalized using a sigmoid activation function to produce the final attention map.
[0035] The network model contains a physical perception module deep down, such as Figure 4 As shown in Figure 2, combining the prior information of the physical model can provide stronger constraints and explanations for the dehazing task. This module is based on the classic atmospheric scattering model:
[0036] In the formula Represents a haze image, represents a clean image without fog, represents the medium transmittance in the atmosphere, Represents the atmospheric light parameters, and the process of reconstructing the defogging recovery result is to use model learning and These two key physical parameters ultimately result in a more accurate fog-free image .
[0037] The input of this module is the deepest layer of the network, and the input is the feature map of the previous level , which contains two branches to reversely deduce and estimate the global atmospheric light and medium transmittance For estimating atmospheric light The branch first passes the global average pooling Learn the overall brightness of the image, and then obtain a feature map with more useful features through dimensionality increase and dimensionality decrease convolution operations, and then use Get the final estimated value. The branch of atmospheric light estimation is different from the branch of atmospheric light estimation in that it does not use global average pooling. Finally, the estimated physical parameters are used to calculate an approximate clear image based on the atmospheric scattering model formula. Through this process, the physical perception module combines physical models with deep learning, providing an effective solution for dehazing tasks.
[0038] Based on the above system, a corresponding image defogging method for low-altitude scenes is obtained. The defogging system to be protected by the present invention is a software system. The data processing logic of the software corresponds to the corresponding method, including the following process: Shoot and acquire low-altitude scene images; The image is input into the above-mentioned low-altitude scene image defogging system for target segmentation, including designing frequency domain fusion modules, space and channel interaction modules, and physical perception modules in the shallow layer, middle layer, and deep layer respectively; The frequency domain fusion module takes as input the three high-frequency components of the feature map processed by wavelet transform. After processing the frequency domain features through a global attention mechanism, it outputs a global frequency domain similarity map, which is used to capture the similarities between pixels in foggy degraded images and use the clearer light fog areas in the image to assist in restoring images in dense fog areas. The input of the spatial and channel interaction module is the feature map after frequency domain fusion. The spatial attention and channel attention of the input feature map are first calculated and concatenated in the channel dimension. Then, random shuffling and group convolution operations are performed to reduce the uneven distribution of haze images at the channel level. The physical perception module is designed based on the atmospheric scattering model. The input is the low-frequency component of the feature map processed by wavelet transform. The low-frequency component information is used as the key parameters of the model. The dehazed and restored image is reconstructed using the atmospheric diffusion model formula in the feature space simulation.
[0039] Based on the data processing flow of each module of the system, the target image after dehazing is output.
[0040] Table 1 Experimental results on two public image dehazing datasets
[0041] Method: method; Pub. / Year: publication conference and year; SOTS-indoor and SOTS-outdoor are both dataset names; PSNR: peak signal-to-noise ratio; SSIM: structural similarity index; Param: parameter size; As shown in Table 1, this example tested the quantitative dehazing performance of the proposed model on two publicly available image dehazing datasets. Compared to the best previous models, the proposed model achieved industry-leading results in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), with significant improvements in PSNR in particular.
[0042] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.
[0043] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. An image defogging system for low-altitude scenes, characterized by: Based on the U-shaped network, the different sensitivities of the U-shaped network to frequency domain information are utilized. By directing different frequency components to different layers, the frequency domain fusion module, space and channel interaction module, and physical perception module are designed in the shallow layer, middle layer, and deep layer respectively. The frequency domain fusion module takes as input the three high-frequency components of the feature map processed by wavelet transform. After being processed by the global attention mechanism of the frequency domain features, it outputs a global frequency domain similarity map, which is used to capture the similarity between pixels in the foggy degraded image and use the clearer light fog areas in the image to assist in restoring the image in the dense fog area. The input of the spatial and channel interaction module is the feature map after frequency domain fusion. The spatial attention and channel attention of the input feature map are first calculated and connected in the channel dimension. Then, random shuffling and group convolution operations are performed to reduce the uneven distribution of haze images at the channel level. The physical perception module is designed based on the atmospheric scattering model. The input is the low-frequency component of the feature map processed by wavelet transform. The low-frequency component information is used as the key parameters of the model. The dehazed and restored image is reconstructed using the atmospheric diffusion model formula in the feature space simulation.
2. The image defogging system for low-altitude scenes according to claim 1, characterized in that: The frequency domain fusion module uses an improved self-attention module to extract features and strengthen the focus on key frequency domain features; The spatial and channel interaction module fuses spatial and channel information from global average pooling and global maximum pooling on a multi-scale structure. The channel information undergoes a random shuffling operation to facilitate cross-channel frequency domain information exchange, while the spatial level is responsible for highlighting areas with significant haze distribution. The physical perception module learns physical parameters through the network, including global atmospheric light and medium transmittance , and combines low-frequency component information to estimate physical parameters and provide physical constraints for the dehazing task.
3. The image defogging system for low-altitude scenes according to claim 1, characterized in that: The frequency domain fusion module is specifically: The encoder part of the network decomposes the input features in the frequency domain through discrete wavelet transform to obtain four frequency domain components. Three high-frequency components, namely the horizontal high-frequency sub-band LH, the vertical high-frequency sub-band HL and the diagonal high-frequency sub-band HH, are selected as the input of the module. The frequency domain fusion module contains three branches. The first branch connects LH, HL and HH, and then performs a connection operation. The second branch uses the horizontal high-frequency texture information LH and the vertical high-frequency texture information for attention calculation. First, the feature dimension is reorganized and the 1×1 convolution is performed to obtain the channel adjustment feature map of the horizontal high-frequency sub-band LH, the vertical high-frequency sub-band HL and the diagonal high-frequency sub-band HH. 、 and , then pass through Normalization obtains an attention matrix of size C×C, where C is the number of channels of the input feature. The third branch uses the HH component to multiply the attention matrix obtained by the second branch to obtain the weighted feature result, and then performs an element-wise addition operation with the residual obtained by the first branch to form a frequency domain fusion feature representation.
4. The image defogging system for low-altitude scenes according to claim 1, characterized in that: The space and channel interaction module is specifically: It contains two branches to learn spatial information and channel information respectively. After merging the two parts in the channel dimension, the channels are randomly shuffled. For the spatial information branch, the feature map output by the previous layer is pooled using global average pooling. and global max pooling After extracting the global features of the spatial dimension, a concatenation operation is performed, followed by a 3×3 convolution operation to learn the spatial attention weights to highlight the areas with significant haze distribution; For the channel information branch, use and After extracting the global features of the channel dimension, a concatenation operation is performed, and then the weights between channels are learned through 1×1 convolution to enhance the modeling of different semantic features; Add the spatial information branch and the channel information branch at the element level, and finally pass the channel shuffle The operation realizes the feature interaction between spatial attention and channel attention; The input feature map is , where the number of channels, height and length of the feature map are C, H and W respectively. The generation process of channel attention first performs the input feature map Perform global average pooling to generate global features in the spatial dimension , using 1×1 convolution Perform feature compression and pass Activation function generates channel attention weight ; First, input feature map Global average pooling and global maximum pooling operations are performed on the channel dimension, the two features are multiplied element by element, and then the spatial attention weights are generated through 3×3 convolution.
5. The image defogging system for low-altitude scenes according to claim 4, characterized in that: The spatial and channel interaction module also introduces a channel shuffling operation to redistribute information between channels, divide the channels into multiple groups, and perform a shuffling operation within each channel group. Through the channel shuffling operation, the rearrangement and interaction of cross-channel information are achieved; the weights after channel shuffling are processed by 3×3 convolution to generate the final channel-specific spatial importance map, and then use The activation function normalizes the importance map to obtain the final attention map.
6. The image defogging system for low-altitude scenes according to claim 2, characterized in that: The formula of the atmospheric diffusion model in the physical perception module is: In the formula Represents a haze image, represents a clean image without fog, represents the medium transmittance in the atmosphere, Represents atmospheric light parameters, and the process of reconstructing the defogging recovery result is to use model learning and These two key physical parameters ultimately result in a more accurate fog-free image ; The input of the module is the deepest layer of the network, and the input is the feature map of the previous level , which contains two branches to reversely deduce and estimate the global atmospheric light and medium transmittance ; For estimating atmospheric light The branch first passes the global average pooling Learn the overall brightness of the image, and then obtain a feature map with more useful features through dimensionality increase and dimensionality decrease convolution operations, and then use Get the final estimated value; finally, use the estimated physical parameters and the atmospheric scattering model formula to calculate the approximate clear image.
7. A method for defogging an image for a low-altitude scene, characterized in that: The following processes are included: Capture and acquire low-altitude scene images; Inputting the image into the low-altitude scene image defogging system according to any one of claims 1 to 6 for target segmentation; Output the target image after dehazing.
Citation Information
Patent Citations
Traffic image defogging method based on non-downsampling contourlet and markov random field
CN108416739A
Unsupervised night image defogging method using high and low frequency decomposition
CN113191964A
Low-light-level image defogging method
CN115496674A
Lightweight image defogging method and device
CN117893444A
Construction method of embedded U-shaped remote sensing defogging network
CN119006845A
Cited By
Remote sensing image haze removal detection method and device and electronic equipment
CN120931629A
Image defogging method based on physical model and prior constraint
CN122222874A
