Image Dehazing Method Based on Physical Fog Sensing and Wavelet Dual-Domain Structure Guidance

By combining wavelet transform and physical fog perception correction module, the problems of insufficient low-frequency fog removal and loss of high-frequency texture information in existing image dehazing methods under complex fog environments are solved, and the image restoration effect is improved while reducing computational complexity.

CN123118370APending Publication Date: 2026-11-17HUAIYIN INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611334120.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing single-image dehazing methods suffer from problems such as insufficient removal of low-frequency haze, loss of high-frequency texture information, and insufficient fusion of physical degradation models and deep features in complex fog environments. Furthermore, the computational complexity of these models is high, making them difficult to adapt to non-uniform fog environments.

Method used

An image dehazing method based on physical fog perception and wavelet dual-domain structure guidance is adopted. The image features are decomposed into low-frequency and high-frequency sub-bands through wavelet transform and processed differently. Combined with the physical fog perception correction module and the cross-frequency structure perception restoration block, the synergistic optimization of fog removal and detail restoration is achieved.

Benefits of technology

It improves the model's adaptability to non-uniform fog environments, reduces computational complexity, improves the structural integrity and visual quality of images, and solves the problems of color shift in low-frequency regions and loss of high-frequency details.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118370A_ABST
    Figure CN123118370A_ABST
Patent Text Reader

Abstract

This invention discloses an image dehazing method based on physical fog perception and wavelet dual-domain structure guidance, belonging to the field of image restoration technology. The method includes: acquiring a preprocessed input foggy image; inputting the input foggy image into an encoder, processing it through multi-level feature encoding modules and downsampling operations to obtain encoded features; inputting the encoded features into a cross-frequency structure-aware restoration block for processing to obtain bottleneck layer features; the cross-frequency structure-aware restoration block includes a wavelet structure-guided restoration module and a scale-adaptive structure selection block; inputting the bottleneck layer features into a decoder, processing it through multi-level upsampling operations, skip connections, and the cross-frequency structure-aware restoration block to output a dehazed image. This invention can effectively improve the adaptability to complex fog environments and enhance the quality of dehazed images. This invention also discloses an electronic device and storage medium for implementing the above method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to image dehazing methods, specifically to an image dehazing method guided by physical fog perception and wavelet dual-domain structure, belonging to the field of image restoration technology. Background Technology

[0002] With the development of intelligent transportation, autonomous driving, smart cities, and intelligent visual perception technologies, image dehazing technology has significant application value in fields such as road monitoring, autonomous driving, and remote sensing detection. However, in actual data acquisition, water vapor and suspended particles in the air can cause light scattering, leading to image degradation such as decreased contrast, color shift, and blurred details, which seriously affects the accuracy of subsequent visual tasks.

[0003] While existing deep learning-based single-image dehazing methods can improve image quality degradation caused by haze to some extent, they still have the following shortcomings in complex scenes:

[0004] First, some existing methods, in order to improve dehazing recovery capabilities, often employ feature enhancement techniques such as increasing network depth, expanding model size, or introducing complex attention mechanisms. This leads to a continuous increase in the number of model parameters and computational complexity, reducing the deployment efficiency of the algorithm on resource-constrained devices. Second, most existing methods directly extract and restore features in the spatial domain, making it difficult to effectively distinguish between haze degradation information and original structural texture information in images. This can easily cause problems such as haze residue in low-frequency areas, color shifts, and loss of high-frequency edge details. Furthermore, in the process of multi-scale feature fusion, existing methods typically use simple feature splicing or element-wise fusion to integrate information at different levels, lacking adaptive interaction mechanisms for different frequency feature characteristics. This can easily introduce redundant information and affect the image structure restoration effect. At the same time, haze in the real environment has obvious spatial non-uniformity. The fog concentration, lighting conditions, and degradation degree vary in different regions, and some existing methods lack the ability to effectively model the fog degradation process, making it difficult to perform adaptive restoration based on the degradation characteristics of different regions. Summary of the Invention

[0005] The purpose of this invention is to provide an image dehazing method based on physical fog perception and wavelet dual-domain structure guidance, to address the problems of insufficient low-frequency fog removal, loss of high-frequency texture information, and insufficient fusion of physical degradation models and deep features in complex fog environments in existing single-image dehazing methods. This invention utilizes wavelet transform to decompose image features into low-frequency degradation information and high-frequency structural information. Through differentiated processing of high and low frequency branches, it achieves synergistic optimization of fog removal and detail restoration. Secondly, a physical fog perception correction module is introduced to model the global fog color features and spatial degradation features in the low-frequency region, improving the network's adaptability to complex fog environments. Furthermore, by combining relational modeling and enhancement mechanisms, the feature restoration capability in the low-frequency region is improved. Simultaneously, high-frequency features are enhanced, and a high-frequency structure guidance mechanism is used to constrain the low-frequency restoration process, reducing texture loss and structural distortion.

[0006] To achieve the above objectives, this invention provides an image dehazing method based on physical fog perception and wavelet dual-domain structure guidance, comprising the following steps:

[0007] Step 1: Obtain the preprocessed input foggy image;

[0008] Step 2: Input the foggy image into the encoder, and process it through multi-level feature encoding modules and downsampling operations to obtain encoded features;

[0009] Step 3: Input the encoded features into the cross-frequency structure-sensing recovery block for processing to obtain the bottleneck layer features;

[0010] Step 4: Input the bottleneck layer features into the decoder, process them through multi-level upsampling operations, skip connections, and cross-frequency structure-aware recovery blocks, and output a dehazed image.

[0011] Furthermore, in step 2, the feature encoding module processes the features, including:

[0012] First, the input features are processed using 1×1 convolution to obtain the mapped features;

[0013] Then, depthwise convolution is used to process the mapped features to obtain spatial local features;

[0014] Subsequently, the local spatial features are input into the SimpleGate gate unit for feature filtering to obtain the filtered features;

[0015] Finally, the filtered features are processed by 1×1 convolution, and then residual fusion is performed with the input features by introducing a learnable scaling parameter to obtain the output features.

[0016] Furthermore, in step 3, the process of processing features in the cross-frequency structure-sensing recovery block includes:

[0017] First, the input features are normalized and then fed into the wavelet structure-guided recovery module for processing to obtain the wavelet structure-guided recovery features.

[0018] Then, the wavelet structure-guided recovery features are residually connected with the input features to obtain the first-stage output features;

[0019] Next, the output features from the first stage are normalized and then input into the scale-adaptive structure selection block for processing to obtain spatial structure compensation features.

[0020] Finally, the spatial structure compensation features are residually connected with the first-stage output features to obtain the second-stage output features.

[0021] Furthermore, the process of the wavelet structure-guided recovery module processing features includes:

[0022] First, the input features are decomposed using discrete wavelet transform to obtain low-frequency sub-band, first high-frequency sub-band, second high-frequency sub-band and third high-frequency sub-band.

[0023] Secondly, after feature mapping of the low-frequency subband, it is input into the physical fog perception correction module for processing to obtain the degradation features and the low-frequency features after physical fog perception correction.

[0024] Then, the first high-frequency sub-band and the second high-frequency sub-band are respectively input to the corresponding high-frequency direction enhancement branches for processing to obtain the enhanced first-direction high-frequency features and the second-direction high-frequency features;

[0025] Next, the high-frequency features in the first and second directions are processed by the orientation calibration structure guidance submodule and then concatenated. After concatenation, they are processed sequentially by 1×1 convolution, depthwise convolution, GELU activation function, 1×1 convolution and Sigmoid function to obtain the structure response map. By introducing 1×1 convolution to perform channel mapping and feature alignment on the structure response map, a high-frequency structure guidance map is obtained.

[0026] Subsequently, the high-frequency structure guidance map, degradation features, and low-frequency features corrected by physical fog perception are input into the low-frequency relationship enhancement module for processing to obtain the enhanced low-frequency features;

[0027] The low-frequency relationship enhancement module's feature processing includes: sequentially processing the low-frequency features corrected for physical fog perception through group normalization, 1×1 convolution, and depthwise convolution to obtain normalized low-frequency features; processing the normalized low-frequency features to generate query features and value features; processing the features concatenated with the degenerate features and normalized low-frequency features to obtain key features; performing element-wise operations on the query features and key features after lightweight convolution mapping to obtain low-frequency relationship responses; performing element-wise operations on the high-frequency structure guidance graph and value features to obtain structure-modulated value features; and performing element-wise operations on the structure-modulated value features and low-frequency relationship responses, followed by residual fusion with the low-frequency features corrected for physical fog perception to obtain enhanced low-frequency features.

[0028] Finally, the enhanced low-frequency features, high-frequency features in the first direction, high-frequency features in the second direction, and high-frequency sub-band in the third direction are spliced ​​together. The spliced ​​features are then processed by inverse discrete wavelet transform to obtain the output features.

[0029] Furthermore, the process of the physical fog perception correction module processing features includes:

[0030] First, the input features are fed into the global haze branch, and then processed sequentially through global average pooling, 1×1 convolution, GELU activation function, and hyperbolic tangent function mapping to obtain the global haze features:

[0031] Secondly, the input features are input into the spatial fog density branch and processed sequentially through depthwise convolution, GELU activation function, 1×1 convolution and Sigmoid function to generate spatial fog density response and apply minimum constraint to it. Then, the constrained spatial fog density response is expanded in the channel dimension to obtain spatial fog density features.

[0032] Then, degradation features are constructed using global fog degradation features and spatial fog density features, and are expressed as follows: , A characteristic of degradation, This is a global fog degradation feature. For element-wise multiplication, Characteristics of spatial fog density;

[0033] Subsequently, the input features and degradation features are concatenated and then input into the degradation correction branch. The concatenated features are processed sequentially through 1×1 convolution, depthwise convolution, nonlinear activation function and 1×1 convolution to generate correction residuals.

[0034] Finally, the corrected residuals are fused with the input features using a learnable scaling factor to obtain the output features.

[0035] Furthermore, the process of enhancing the branching feature in the high-frequency direction includes:

[0036] First, the input features are processed by 1×1 convolution and then fed into a dual-branch structure for processing. One branch generates directional structure features through directional convolution, while the other branch generates local texture features through depth convolution, SimpleGate gate unit and 1×1 convolution in sequence.

[0037] Then, the directional structural features and local texture features are concatenated and fused using a 1×1 convolution to obtain intermediate features;

[0038] Next, the intermediate features after gating are multiplied by the learnable scaling factor, and then added to the input features through residual connection to obtain the output features.

[0039] Furthermore, the process of the orientation calibration structure guidance submodule processing features includes:

[0040] The input features are processed sequentially through 1×1 convolution and directional convolution to obtain initial directional structure features; then, the initial directional structure features are processed using depthwise convolution to obtain directional structure features.

[0041] The directional structural features are input into the channel-gated branch and processed by global average pooling, 1×1 convolution, GELU activation function, 1×1 convolution and Sigmoid function to obtain the channel-gated weights;

[0042] The directional structural features are input into the sub-band spatial gating branch, and after being processed by channel-dimensional average pooling and channel-dimensional max pooling respectively, feature concatenation, 7×7 convolution and Sigmoid function processing are performed sequentially to obtain the sub-band spatial gating weights.

[0043] The initial directional structural features are weighted element-wise using channel gating weights and subband spatial gating weights, and then channel fusion is performed using 1×1 convolution to obtain weighted directional structural features.

[0044] The weighted directional structural features are added to the input features by residual addition to obtain the output directional structural features.

[0045] Furthermore, the process of processing features by the scale-adaptive structure selection block includes:

[0046] First, multi-scale feature extraction is performed on the input features after 1×1 convolution, and the extracted multi-scale structural features are concatenated to obtain the concatenated features.

[0047] Then, the concatenated features are processed sequentially through global average pooling, 1×1 convolution, non-linear activation function, 1×1 convolution and Sigmoid function to obtain scale-adaptive weights;

[0048] Next, the multi-scale structural features are weighted using scale-adaptive weights, and the weighted multi-scale structural features are then processed by feature concatenation, global average pooling, and 1×1 convolution to obtain the compensated features.

[0049] Finally, the compensated features and input features are residually fused to obtain the output features.

[0050] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0051] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] 1. This invention uses a physical fog perception correction module to jointly model global fog degradation features and spatial degradation features. Global fog degradation features reflect the overall fog color shift and brightness degradation trend of an image, while spatial degradation features reflect the differences in fog concentration and degradation degree in different regions. By using both for feature correction, the network can adaptively adjust the restoration intensity according to the fog distribution in different regions of the image, avoiding the problem of insufficient or excessive local defogging caused by applying a uniform restoration strategy to the entire image. Therefore, this invention can improve the model's adaptability to non-uniform fog environments.

[0054] 2. This invention decomposes image features into low-frequency and high-frequency sub-bands using wavelet transform. The low-frequency sub-band mainly contains brightness, color, and overall structural information, while the high-frequency sub-band mainly contains edge, texture, and local detail information. Since haze degradation typically causes color shift and contrast reduction in low-frequency regions, as well as edge blurring and texture weakening in high-frequency regions, directly processing these uniformly in the spatial domain can easily lead to insufficient low-frequency recovery or loss of high-frequency details. This invention employs different processing strategies for low-frequency and high-frequency information, focusing on restoring brightness and color degradation in the low-frequency branch and enhancing edge and texture information in the high-frequency branch. This improves the targeting of dehazing restoration and reduces structural loss and over-smoothing.

[0055] 3. This invention incorporates physical degradation information into the low-frequency feature relationship modeling process through a low-frequency relationship enhancement module. Physical degradation information characterizes the residual haze area and degradation intensity. By integrating this information into the low-frequency relationship modeling, the network can focus more on severely degraded haze areas when calculating low-frequency feature responses, thereby improving the ability to extract effective low-frequency information. Simultaneously, the enhanced high-frequency structural features are used to compensate for detail information in the low-frequency recovery process, ensuring that low-frequency recovery is not only dependent on color and brightness information but also constrained by structural information such as edges and textures. Therefore, this invention can improve problems such as edge blurring, texture loss, and contour distortion, enhancing the structural integrity and visual quality of the recovered image.

[0056] 4. This invention reduces the computational complexity of the model by using a lightweight feature encoding module and a deep convolutional structure. The lightweight feature encoding module can extract the basic spatial features of the input image with fewer parameters, while the deep convolutional structure reduces the parameter and computational overhead of ordinary convolution by extracting local features by channel. Furthermore, by combining frequency domain differential processing after wavelet decomposition, the network can reduce computational complexity while maintaining dehazing effects. Therefore, this invention is better suited for fast image restoration tasks in real-world visual perception scenarios. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall structure of the defogging network provided by the present invention;

[0058] Figure 2 This is a schematic diagram of the feature encoding module structure provided by the present invention;

[0059] Figure 3 This is a schematic diagram of the wavelet structure-guided recovery module provided by the present invention;

[0060] Figure 4 This is a schematic diagram of the orientation calibration structure guidance submodule provided by the present invention;

[0061] Figure 5 This is a schematic diagram of the physical fog sensing and correction module provided by the present invention;

[0062] Figure 6 A schematic diagram of the scale-adaptive structure selection block structure provided by the present invention;

[0063] Figure 7 This is a schematic diagram of the defogging result provided by the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0065] Example 1:

[0066] To address the problems of reduced contrast, color shift, low-frequency haze residue, and loss of high-frequency texture details in images under complex fog environments, this invention proposes an image dehazing method based on physical fog perception and wavelet dual-domain structure guidance.

[0067] like Figure 1 As shown, the dehazing network adopts an encoder-bottleneck-decoder structure. The encoder mainly consists of a shallow feature extraction layer, a lightweight feature encoding module (LFEM), and a downsampling layer, used to extract image features at different scales step by step. The bottleneck layer contains a cross-frequency structure-aware restoration block (CFSARB), used to collaboratively restore low-frequency fog degradation information and high-frequency structure information from deep features. The decoder consists of an upsampling layer and a cross-frequency structure-aware restoration block, used to progressively restore the feature spatial resolution and continue cross-frequency structure-aware restoration at the corresponding scale. Finally, the network generates a dehazed image through an image reconstruction layer. The LFEM module, located in the network encoding stage, extracts features from the input hazy image step by step through a convolutional structure, obtaining effective spatial feature representations while reducing model computational complexity, and forming semantic features at different scales through multi-scale downsampling operations.

[0068] In the network bottleneck layer and decoding stage, a CFSARB module is introduced. This module uses discrete wavelet transform to decompose the input features into low-frequency sub-bands and multiple high-frequency sub-bands, and performs targeted processing based on the characteristics of different frequency features. The low-frequency components are mainly used to describe information such as haze degradation, color shift, and brightness changes, while the high-frequency components are used to characterize edge, texture, and structural details. By constraining the low-frequency recovery process with high-frequency structural information, collaborative recovery of frequency features is achieved.

[0069] Finally, after step-by-step upsampling by the decoder and fusion of multi-scale features through skip connections in the encoding stage, the image resolution is restored to the original resolution, and the final dehazing result is output through a convolutional layer. The specific implementation steps are as follows:

[0070] Step 1: Obtain the publicly available Haze-4k dataset. The Haze-4k dataset was proposed by Liu et al. in their ACM MM 2021 paper, "From Synthetic to Real: Image Dehazing Collaborating with Unlabeled RealData". This dataset contains 4000 images of mixed indoor and outdoor scenes. Following the official Haze-4k dataset partitioning method, 3000 images were used as the training set and 1000 images as the test set to ensure the reliability of the evaluation.

[0071] The acquired Haze-4k dataset was preprocessed. Specifically, this included: first, pairing and checking the foggy images with their corresponding clear images in the dataset to ensure consistency in file name, image size, and scene content; then, converting the images to a uniform RGB three-channel format and normalizing the pixel values ​​to the range of 0-1 to obtain normalized foggy and clear images.

[0072] Step 2: Input the preprocessed foggy image into the encoder, and it will pass through shallow feature mapping, two-level lightweight feature encoding modules and two downsampling operations in sequence to obtain deep encoded features.

[0073] Specifically, the preprocessed foggy image obtained in step 1 is denoted as... ,in The size is B represents the batch size, and H and W represent the height and width of the image, respectively.

[0074] First, a 3×3 convolution is applied to the input foggy image. Shallow feature mapping is performed to map the input image from the RGB image space to the feature space, thus obtaining the initial features. , represented as: ,

[0075] in, express Convolution operations are used to extract basic color, edge, and local texture information from an input image.

[0076] Subsequently, the initial features Inputting into the first-level feature encoding module yields the first-level encoded features. , represented as:

[0077] Where LFEM represents the feature encoding module. According to... Figure 2 The structure is as follows: the feature encoding module includes a main branch and a residual branch. For the input features... The main branch includes a 1×1 convolution, a 3×3 depthwise convolution, a SimpleGate gate unit, and a second 1×1 convolution. Its output is then added to the input features as a residual after passing through a learnable scaling parameter β.

[0078] Specifically, the input features are first processed through a first 1×1 convolution. Perform channel mapping and inter-channel information exchange to obtain features. : ,

[0079] in, This represents a 1×1 convolution at the front end of the feature encoding module, used to adjust feature channels and fuse channel information. Then, Input a 3×3 depthwise convolution to extract local spatial structure information and obtain features. :

[0080] ,

[0081] in, This represents a 3×3 depthwise convolution operation, used to extract local edges and texture structures with low computational complexity. Subsequently, the features... Input the SimpleGate gate unit for feature filtering to obtain features. :

[0082] ,

[0083] The SimpleGate gate is used to adaptively filter effective responses from input features, thereby enhancing feature information related to edges, color variations, and local textures.

[0084] Next, channel fusion is performed on the gated features through a second 1×1 convolution to obtain the main branch output features. :

[0085] ,

[0086] in, This represents a 1×1 convolution at the back end of the feature encoding module, used to complete channel fusion and feature remapping.

[0087] Finally, output the features of the main branch. After adjustment by the learnable scaling parameter β, and compared with the input features By summing the residuals, we obtain the output feature Y of the feature encoding module: ,

[0088] Here, β is a learnable scaling parameter used to control the strength of the influence of the main branch's enhanced features on the input features. This residual fusion method can enhance local structural representation while preserving input features and improve the stability of the network training process.

[0089] Therefore, the first-level feature encoding module can generate initial features Mapped to first-level encoded features It is used to preserve shallow edges, color variations, and local texture information in the input image.

[0090] Furthermore, the first-level encoded features are downsampled to obtain intermediate features with reduced spatial scale and increased channel dimension. , represented as: ,

[0091] Here, Down represents the downsampling operation, which is used to reduce the spatial resolution of the feature map, expand the receptive field of the feature, and obtain a higher-dimensional semantic feature representation.

[0092] Subsequently, the intermediate features are input into the second-level feature encoding module to further extract deep local structural features and semantic features, thus obtaining the second-level encoded features. , represented as: ,

[0093] Next, the second-level encoded feature E2 is downsampled again to obtain the bottleneck layer input feature. , represented as:

[0094] ,

[0095] Among them, the bottleneck layer input features It has a larger receptive field and higher semantic expression capability, which is used as input to the subsequent wavelet dual-domain structure-guided recovery module for frequency decomposition, low-frequency fog degradation recovery and high-frequency structure enhancement processing.

[0096] Simultaneously, the first-level and second-level encoded features are used as skip connection features, respectively, and passed to the feature fusion positions at the corresponding scales in the subsequent decoding stage to supplement the shallow spatial details and local structural information that may be lost during the decoding process. Through the above two-level encoding process, this invention can obtain multi-scale encoded features containing color, edge, texture, and local structural information while controlling the computational complexity of the model, providing a stable feature foundation for subsequent wavelet dual-domain structure-guided recovery.

[0097] Step 3: Input the bottleneck layer features obtained in Step 2. The input is fed into the cross-frequency structure sensing recovery block for frequency domain recovery and spatial structure compensation to obtain the bottleneck layer recovery features.

[0098] Specifically, firstly, input features are applied to the bottleneck layer. Normalization is performed to obtain normalized bottleneck characteristics. ; then, The input wavelet structure-guided restoration module decomposes the features into low-frequency and high-frequency sub-bands using discrete wavelet transform, and then performs differential restoration on low-frequency fog degradation information, high-frequency directional structure information, and high-frequency texture detail information, respectively. Next, the restored low-frequency and high-frequency sub-bands are re-fused using inverse discrete wavelet transform to obtain the bottleneck layer frequency structure restoration features. Then, Input features of the bottleneck layer By summing the residuals, we obtain the bottleneck recovery characteristics of the first stage. Finally, The input scale-adaptive structure selection block compensates for potential local texture loss, edge discontinuities, and insufficient spatial structure restoration after wavelet restoration, resulting in bottleneck layer restoration features. : ,

[0099] like Figure 3 and Figure 6 As shown, the CFSARB module adopts a Transformer-like residual structure, mainly including a Wavelet Structure-Guided Restoration Module (WSGRM) and a Scale-Adaptive Structure Selection Block (SASSB). WSGRM is used for wavelet frequency decomposition, high-low frequency differential restoration, and high-frequency structure-guided low-frequency restoration of the input features; SASSB is used for multi-scale spatial structure compensation of the wavelet-restored features.

[0100] For bottleneck layer input features First, normalization is performed to obtain the normalized bottleneck characteristics. :

[0101] ,

[0102] Norm represents the normalization operation, which is used to stabilize the feature distribution and improve the training stability of the subsequent wavelet structure-guided recovery module.

[0103] Then normalize the bottleneck features Input WSGRM, perform frequency decomposition and high / low frequency structure recovery to obtain the bottleneck layer frequency structure recovery features. : ,

[0104] in, This represents the bottleneck layer frequency structure recovery feature output by WSGRM. This feature includes fog degradation, brightness, and color information recovered from the low-frequency branch, as well as edge, texture, and orientation structure information enhanced from the high-frequency branch. The process of constructing the wavelet structure-guided recovery module is as follows:

[0105] First, the input features are decomposed using Discrete Wavelet Transform (DWT):

[0106] ,

[0107] in: This is the low-frequency sub-band, which mainly contains information on overall image brightness, color, and haze degradation. For horizontal high-frequency sub-band, For vertical high-frequency sub-bands; It includes local texture and detail information.

[0108] For the high-frequency subband HH portion, this invention retains the original features and does not perform any additional enhancement processing. For the directional high-frequency subband... and The present invention adopts respectively Convolution pairs of high-frequency directional subbands and Perform channel mapping.

[0109] ,

[0110] ,

[0111] in, and Represent and go through Features of convolutional channel mapping This represents a 1×1 convolution operation.

[0112] For the high-frequency sub-bands in the horizontal direction, horizontal strip depth convolution is used to extract the horizontal edge response:

[0113] ,

[0114] in, represent The high-frequency structural features in the horizontal direction obtained after 1×k convolution mapping This represents a horizontal stripe depth convolution operation with a kernel size of 1×k, used to enhance horizontal edges and textures. K represents the length of the stripe convolution sum, and in this experiment, k is set to 5.

[0115] For the high-frequency subbands in the vertical direction, vertical strip depth convolution is used to extract the vertical structural response:

[0116] ,

[0117] in, express The high-frequency structural features in the vertical direction obtained after k×1 convolution mapping This represents a vertical stripe depthwise convolution operation with a kernel size of k×1, used to enhance vertical edges and textures. K represents the length of the stripe convolution sum, and in this experiment, k is set to 5.

[0118] In addition, to supplement local texture information, a local texture enhancement branch is introduced:

[0119] ,

[0120] in, and These represent the local texture enhancement features extracted from high-frequency embedding features in the corresponding directions; This represents a 3×3 depth convolution operation; This represents a gating unit used to filter valid high-frequency texture responses; This represents a 1×1 convolution operation.

[0121] Subsequently, the directional structural features and local texture features are concatenated, and then subjected to 1×1 convolution, 3×3 depthwise convolution, and 1×1 convolution to obtain the enhanced directional high-frequency features:

[0122] ,

[0123] in, and To obtain new high-frequency features in the horizontal and vertical directions, These are learnable scaling factors used to control the high-frequency enhancement intensity. This represents a 1×1 convolution operation. This represents a 3×3 convolution operation. This represents the sigmoid function. This indicates a channel-level concatenation operation. and This represents the high-frequency boosting weights generated by a 3×3 convolution and the sigmoid function. represent The high-frequency structural features in the horizontal direction obtained after 1×k convolution mapping express The high-frequency structural features in the vertical direction obtained after k×1 convolution mapping and These represent the local texture enhancement features extracted from the high-frequency embedding features in the corresponding directions.

[0124] To further improve the reliability of high-frequency structure guidance maps, this invention generates high-frequency structure guidance maps... Previously, introduced Figure 4 The orientation calibration structure guidance submodule shown performs orientation structure extraction, channel selection, and spatial location selection on the enhanced orientation high-frequency features. This submodule functions as follows: Figure 3 Enhanced horizontal high-frequency characteristics of the mid-to-high frequency direction enhancement branch output and enhance high-frequency features in the vertical direction .

[0125] Specifically, for high-frequency features in the input direction ,in It can represent the enhanced horizontal high-frequency features. It can also represent the enhanced high-frequency features in the vertical direction. First, the high-frequency features of the input direction are... The initial directional structural features are obtained by sequentially performing 1×1 convolution and directional strip depthwise convolution. Then, through 3×3 depthwise convolution, enhanced directional structural features are obtained. :

[0126] ,

[0127] ,

[0128] in, This represents a 1×1 convolution operation; This indicates a directional stripe depthwise convolution operation, used to extract directional structural responses from high-frequency directional features; This indicates a 3×3 depthwise convolution operation, used to further enhance the continuity of local edges, textures, and orientation structures; Represents the initial directional structural features. This indicates the enhanced directional structural features.

[0129] When the input feature X is an enhanced horizontal high-frequency feature When the input features are used, the directional strip depthwise convolution uses a 1×5 horizontal directional strip depthwise convolution; For enhanced vertical high-frequency features At this time, the directional strip depth convolution uses a 5×1 vertical directional strip depth convolution. Therefore, the directional structure enhancement processes in the two directions can be represented as follows:

[0130] ,

[0131] ,

[0132] in, This indicates the enhanced horizontal high-frequency features. Extracted enhanced directional structural features; This indicates the enhanced high-frequency features in the vertical direction. The extracted enhanced directional structural features.

[0133] Subsequently, the directional structural features will be enhanced. By inputting the channel gating branch and the spatial gating branch respectively, the channel gating weights are obtained. and spatial gating weights .

[0134] In the channel gating branch, for Global average pooling is performed to obtain channel statistical features; then, the channel gating weights are generated by sequentially passing them through 1×1 convolution, GELU nonlinear activation function, 1×1 convolution, and Sigmoid function. :

[0135] ,

[0136] in, This represents the channel gating weight, used to filter valid directional structural responses from the channel dimension; This indicates a global average pooling operation; and These represent the two 1×1 convolution operations in the channel-gated branch, respectively. This represents the GELU nonlinear activation function; This represents the sigmoid function.

[0137] In the spatial gating branch, respectively for Average pooling and max pooling are performed along the channel dimension to obtain the spatial average response and spatial maximum response; then the two are concatenated and successively passed through a 7×7 convolution and a sigmoid function to generate spatially gated weights. : ,

[0138] in, Represents spatial gating weights, used to filter reliable edge, texture, and structural responses in the positional dimension of the high-frequency subband feature map; This indicates that max pooling is performed along the channel dimension; This indicates that average pooling is performed along the channel dimension; This represents a 7×7 convolution operation.

[0139] Then, enhance the directional structural features Channel gating weights and spatial gating weights Element-wise weighting is performed, and channel fusion is achieved through 1×1 convolution to obtain orientation calibration features. : ,

[0140] in, This represents element-wise multiplication; This indicates the directional structural features after joint calibration by channel gating and space gating. Represents the channel gating weight. Represents spatial gating weights.

[0141] Finally, the orientation calibration feature High-frequency characteristics of the input direction By summing the residuals, the output characteristics of the orientation calibration structure guidance submodule are obtained. :

[0142] ,

[0143] in, This represents the high-frequency structural characteristics after direction calibration, where X represents the high-frequency characteristics in the input direction. This represents the orientation calibration feature. Through this residual connection, reliable orientation edge, texture, and structural responses can be further enhanced while suppressing invalid or noisy high-frequency information, while preserving the high-frequency characteristics of the input orientation.

[0144] Therefore, for the enhanced horizontal high-frequency features and enhanced vertical high-frequency features After passing through Figure 4 After processing by the orientation calibration structure guidance submodule, the high-frequency structural features after orientation calibration are obtained:

[0145] ,

[0146] ,

[0147] in, This indicates the high-frequency structural features in the horizontal direction after orientation calibration. This indicates the high-frequency structural features in the vertical direction after orientation calibration. and Represents the initial directional structural features in the horizontal and vertical directions. and These represent the channel gating weights and spatial gating weights corresponding to the high-frequency features in the horizontal direction, respectively. and These represent the channel gating weights and spatial gating weights corresponding to the high-frequency features in the vertical direction, respectively.

[0148] Subsequently, the high-frequency structural features after direction calibration are stitched together along the channel dimension to obtain the high-frequency structural fusion features. :

[0149] ,

[0150] in, This indicates the enhanced horizontal high-frequency features. and vertical high frequency features High-frequency structural fusion features obtained by splicing.

[0151] Subsequently, structural information is fused using 1×1 convolution, depthwise convolution, and 1×1 convolution, and a high-frequency structural guidance map is generated using Sigmoid activation. :

[0152] ,

[0153] in, This represents the guiding weights generated by the directional high-frequency structure, used to modulate the value feature V during the low-frequency recovery process. This represents the sigmoid function. This represents a 3×3 depthwise convolution. This represents a 1×1 convolution.

[0154] In low-frequency subband First, feature mapping is performed on the low-frequency sub-band: ,

[0155] Then The input to the Physics-aware Haze Attention Module (PHAAM) yields low-frequency features corrected for physical fog perception. .

[0156] like Figure 5 PHAAM includes a global fog color branch, a spatial fog density branch, and a degradation correction branch, used to construct a fog degradation representation in the feature space and perform physical perception correction on low-frequency features. The specific process of constructing the physical fog perception correction module is as follows:

[0157] First, the global fog branch obtains the overall statistical information of low-frequency features through global average pooling (GAP), and then obtains the global fog degradation feature A through 1×1 convolution, non-linear activation function, and hyperbolic tangent function mapping.

[0158] ,

[0159] In this context, feature A represents the global fog degradation feature. It is used to describe the overall haze shift and global degradation trend in low-frequency features. denoted by GELU, it represents a nonlinear activation function, and tanh represents the hyperbolic tangent function. B represents a 1×1 convolution, B represents the batch size, and C represents the number of feature channels.

[0160] Secondly, the spatial fog density branch extracts the spatial degradation distribution through a 5×5 depthwise convolution and generates the spatial fog density response through a 1×1 convolution and a sigmoid function. :

[0161] ,

[0162] in, This represents the sigmoid function, which restricts the output to between 0 and 1. Represents a 1×1 convolution. This represents a 5×5 depthwise convolution.

[0163] To avoid excessive enhancement of the low-frequency recovery process due to an excessively small spatial fog density response, Apply a minimum constraint:

[0164] ,

[0165] in, This is a preset lower limit parameter, typically ranging from 0 to 1. When the value is small, the model has a higher degree of freedom in correcting local haze degradation, but this may increase the risk of local over-enhancement or color instability; when... When the value is larger, the low-frequency recovery process is more stable, but it may weaken the model's adaptive recovery ability for areas with slight haze. Therefore, this embodiment will... Set it to 0.05 to strike a balance between recovery strength and training stability.

[0166] Then Extending the channel dimension yields spatial fog density features consistent with the number of low-frequency feature channels: , , For broadcast extension, it is a general tensor operation rule, where B is the batch size, C represents the number of feature channels, H represents the height, and W represents the width.

[0167] Based on the fog degradation relationship in the feature space, degradation features are constructed using global fog degradation feature A and spatial fog density feature T: , ⊙ represents element-wise multiplication. , is used to represent the degradation component formed by the combination of global fog color information and spatial fog density.

[0168] Subsequently, the original low-frequency features The degradation feature D and the spatial fog density feature T are concatenated, and the degradation correction branch is used to generate the correction residual. : ,

[0169] in, The low-frequency correction residual generated by the degradation correction branch is used to compensate for the residual haze degradation, brightness attenuation and color shift in the low-frequency features; D represents the degradation feature jointly constructed by the global haze color feature and the spatial haze density feature, and T represents the spatial haze density feature.

[0170] Finally, through learnable scaling factors The intensity of the correction residual is controlled and fused with the input low-frequency features to obtain the low-frequency features after physical fog perception correction: , This indicates the low-frequency characteristics after physical fog perception correction; This represents a learnable scaling factor used for adaptive control to correct residuals. For low-frequency features The impact intensity. This residual fusion method can introduce haze correction information generated by the degradation correction branch while preserving the original low-frequency content information.

[0171] Through the above process, PHAAM can explicitly model the global fog color shift, spatial fog density variation, and degradation features formed by the two in low-frequency characteristics, giving the low-frequency recovery process stronger physical constraints and adaptability to non-uniform fog.

[0172] After passing through the Physical Fog Sensing Correction (PHAAM) module, the low-frequency characteristics after physical fog sensing correction are obtained. Simultaneously, degradation features D, used to characterize haze degradation components, are obtained. To further enhance the identification and structural recovery capabilities of haze residues in low-frequency regions, this invention utilizes the output features of PHAAM. After group normalization, convolution, Unified low-frequency features are obtained after depthwise convolution processing, before low-frequency relationship modeling. ,

[0173] ,

[0174] in, This indicates a group normalization operation used to stabilize the distribution of low-frequency features; Represents a 1×1 convolution; This represents a 3×3 depthwise convolution; This represents the unified low-frequency features after normalization and convolutional reintegration.

[0175] The low-frequency relationship enhancement branch is then used for Q, K, and V feature modeling. Unlike ordinary attention mechanisms where query feature Q, key feature K, and value feature V are all directly generated from the same feature, this invention introduces the degradation feature D obtained from PHAAM into feature K and the high-frequency structure guidance map Gs into feature V. This approach enables low-frequency relationship modeling to simultaneously possess fog degradation perception capabilities and high-frequency structure constraint capabilities.

[0176] Specifically, from the corrected low-frequency characteristics Generate query feature Q and value feature V:

[0177] , ,

[0178] Where Q represents the response of the region to be recovered in the low-frequency features, and V is used to retain low-frequency content information.

[0179] For feature K, this invention will use low-frequency features Feature K is obtained by fusing it with the degenerate feature D:

[0180] ,

[0181] Where D represents the degradation feature generated by PHAAM.

[0182] Through this design, the K feature not only contains low-frequency image content information, but also degradation prior information composed of global fog degradation features and spatial fog density features, thus making subsequent relation modeling more focused on the residual fog and haze areas.

[0183] Subsequently, convolutional mappings were performed on the query feature Q and the key feature K respectively:

[0184] For query feature Q, given low-frequency features Directly generated, going through the following steps convolution, Depth convolution, non-linear activation function GELU and Convolution processing yields query features after convolution mapping. :

[0185] ,

[0186] in, This represents the query features after convolutional mapping. and These represent the front and back 1×1 convolutions of feature Q, respectively. This represents a 3×3 depthwise convolution.

[0187] For feature K, it is composed of low-frequency features. The degradation-aware key feature is obtained by concatenating the degradation feature D generated by PHAAM along the channel dimension, and then sequentially processing it through 1×1 convolution, 3×3 depthwise convolution, GELU nonlinear activation function, and 1×1 convolution. : ,

[0188] in, This represents the features of the degradation-sensing key. and The representative feature K is formed by two 1×1 convolutions. This represents a 3×3 depthwise convolution.

[0189] For feature V, derived from low-frequency features The feature V is generated directly without convolution mapping and is retained.

[0190] right and To model low-frequency relationships, we first calculate the relationship response between query features and degenerate-aware key features through element-wise correlation operations: .

[0191] in, and represents the query feature and key feature after convolution mapping, respectively; R represents the relational response at different spatial locations in the low-frequency region; and ⊙ represents element-wise multiplication.

[0192] Simultaneously, the high-frequency structure-guided map Gs is applied to the value feature V to obtain the structure-guided modulated feature Vg: ,

[0193] in, This represents the value characteristics after high-frequency structure-guided modulation. In this way, Vg not only retains low-frequency content information but also adaptively enhances based on high-frequency edges and texture structures. ⊙ represents element-wise multiplication.

[0194] Subsequently, the low-frequency relational response R is fused with the value feature Vg after high-frequency structure-guided modulation: ,

[0195] in, The characteristic Vg represents the fusion of the low-frequency response R and the high-frequency structure-guided modulation value.

[0196] This process adaptively adjusts the recovery region using degradation-aware relational responses, while simultaneously constraining low-frequency content recovery with high-frequency structural information, enabling the network to remove haze while preserving edge continuity and texture details. Finally, the initial low-frequency sub-band features are... Low-frequency characteristics after physical fog perception correction Perform residual fusion: ,

[0197] in, Represents the initial low-frequency subband characteristics Low-frequency characteristics after physical fog perception correction Features after residual fusion.

[0198] Through the above process, the degradation information guides the K branch and the high-frequency structural information constrains the V branch. Specifically, the degradation feature D improves the ability of low-frequency relation modeling to identify hazy areas, and the high-frequency structural guidance map Gs enhances the edge and texture representation capabilities in the restoration results, thereby improving the ability to restore image brightness, correct color, and preserve structural details in complex fog environments.

[0199] After completing the low-frequency subband recovery, high-frequency directional enhancement, and high-frequency structure guidance, the recovered low-frequency subband... Enhanced horizontal high-frequency subband Enhanced vertical high-frequency subband and diagonal high-frequency subband By inputting the inverse discrete wavelet transform, low-frequency content information and high-frequency structural information are re-fused to obtain the bottleneck layer frequency structure recovery features of the WSGRM output. : ,

[0200] Wherein, IDWT represents the inverse discrete wavelet transform operation, which is used to reconstruct the recovered low-frequency subband and high-frequency subband into frequency structure recovery features with the same size as the input feature space. This indicates the low-frequency output characteristics after physical fog perception correction and low-frequency relationship enhancement; and These represent the enhanced high-frequency features in the horizontal and vertical directions, respectively. Indicates diagonal high-frequency sub-bands; This represents the bottleneck layer frequency structure recovery feature output by the wavelet structure-guided recovery module.

[0201] Then, the frequency structure of the bottleneck layer is restored to its characteristic. The bottleneck layer input features obtained in step 2 By summing the residuals, we obtain the bottleneck recovery characteristics of the first stage. : ,

[0202] in, This represents the bottleneck features of the first stage after wavelet dual-domain structure recovery. The residual connection is used to preserve the original semantic features extracted by the lightweight coding module in step 2. At the same time, frequency domain recovery information generated by WSGRM is introduced to avoid excessive perturbation of the original coded features during the frequency domain processing.

[0203] Furthermore, the bottleneck recovery features from the first stage are normalized again and input into SASSB to obtain the bottleneck layer scale-adaptive structural compensation features. : ,

[0204] in, This represents the multi-scale spatial structure compensation feature output by SASSB. This feature is used to adaptively select local, mesoscale, and large-scale structural information based on the structural complexity and fog degradation degree of different regions, thereby compensating for problems such as local texture loss, edge discontinuity, and insufficient spatial structure recovery that may exist after wavelet restoration.

[0205] Furthermore, the process of constructing the scale-adaptive structure selection block is as follows:

[0206] After processing by Cross-Frequency Structure-Aware Reconstruction Block (CFSARB), the features have completed the information recovery of different frequency subbands. However, since wavelet transform mainly decomposes and reconstructs frequency information, it still has shortcomings in addressing issues such as spatial structure discontinuities, local texture loss, and differences in degradation levels in different regions that remain after the recovery process.

[0207] To address the aforementioned issues, this invention proposes a Scale-Adaptive Structure Selection Block (SASSB). By constructing multi-scale spatial feature branches with different receptive fields and dynamically adjusting the importance of features at different scales using a scale-adaptive selection mechanism, the residual structural information after wavelet recovery is further optimized.

[0208] For the output features of the first stage First, channel mapping is performed using a 1×1 convolution:

[0209] ,

[0210] in, This indicates the spatial characteristics after the channel has been expanded.

[0211] To obtain structural information at different spatial scales, this invention constructs three feature extraction branches with different receptive fields. The local scale branch uses a 3×3 depthwise convolution: , Indicates local scale structural features; This indicates a depthwise convolution operation with a kernel size of 3×3, used to extract local edge, texture, and detail structure information; This represents the input features after 1×1 convolution channel mapping.

[0212] The mesoscale branch uses a 5×5 depthwise convolution: ,

[0213] in, Indicates mesoscale structural features; This indicates a depthwise convolution operation with a kernel size of 5×5, used to extract edge continuity, region texture, and contextual structure information within a moderate receptive field.

[0214] Large-scale branches employ dilated depthwise convolution: ,

[0215] in, Represents large-scale structural features; d represents a 3×3 depthwise convolution operation with a dilation rate of 2, used to expand the receptive field without significantly increasing the number of parameters, thereby extracting a wider range of structural contours and spatial context information; d represents the dilation rate.

[0216] The three scale branches extract recovery features from different spatial ranges, providing multi-scale candidate information for subsequent structure selection.

[0217] Because the degradation levels and structural characteristics vary across different foggy environments, fixed-scale fusion methods cannot guarantee optimal restoration results for all regions. Therefore, this invention utilizes multi-scale features to jointly generate scale selection weights. First, the three scale features are fused: ,

[0218] in, Indicates local scale features Mesoscale features and large-scale features Multi-scale fusion features obtained by splicing.

[0219] Subsequently, a scale response is generated using global context information: ,

[0220] in, Indicates multi-scale fusion features The extracted global context description vector; This represents a global average pooling operation, used to aggregate global statistical information on structural features at different scales.

[0221] The scale weights are obtained through nonlinear mapping using two layers of 1×1 convolution kernels: ,

[0222] in, , representing the importance weights of the three scale branches respectively: Select weights for the local structure; Select weights for the regional context; To select weights for a large-scale structure, and This represents two consecutive 1×1 convolutions.

[0223] In this way, the network can adaptively select features of different scales to participate in the restoration process based on the spatial degradation of the input image.

[0224] The generated scale weights are used to dynamically modulate the features at the three scales respectively:

[0225] ,

[0226] in, , and These represent the structural features output by the local scale branch, the mesoscale branch, and the large scale branch, respectively. , and These represent the scale selection weights corresponding to the three scale branches; , and These represent the weighted local-scale structural features, mesoscale structural features, and large-scale structural features, respectively.

[0227] Subsequently, the weighted multi-scale structural features are fused: ,

[0228] in, This indicates a channel-level concatenation operation; express Convolution operations are used to perform channel fusion and channel compression on the spliced ​​multi-scale structural features; The initial structural compensation features output by the scale-adaptive structural selection block are used to compensate for potential issues such as local texture loss, edge discontinuities, and insufficient spatial structure recovery after wavelet restoration. Channel compression is achieved by extracting global contextual statistics through global pooling and performing 1×1 convolution.

[0229] ,

[0230] in, This represents the scale-adaptive structural compensation feature of the bottleneck layer. This represents the initial structural compensation features output by the scale-adaptive structural selection block. Represents global pooling. express Convolution operation.

[0231] Through the above process, SASSB can adaptively select local, mesoscale, and large-scale information based on the structural characteristics of different regions to achieve spatial structure compensation after wavelet restoration. Compared with fixed multi-scale fusion methods, this module can effectively avoid feature redundancy caused by the introduction of invalid scale information, and improve the structural continuity and texture preservation ability of the restoration results.

[0232] Finally, the bottleneck layer scale adaptive structural compensation feature is applied. Characteristics of bottleneck recovery in the first stage By summing the residuals, the output features of the bottleneck layer are obtained. : ,Right now: ,

[0233] in, The output features of the bottleneck layer CFSARB simultaneously include the original semantic information obtained by lightweight coding in step 2, wavelet frequency domain recovery information, and multi-scale spatial structure compensation information.

[0234] Step 4: Recover the bottleneck layer features obtained in Step 3 The input is processed by the decoder, which restores the feature space resolution step by step to obtain the decoded and restored features.

[0235] Furthermore, the process of building the decoder is as follows:

[0236] After processing by the encoder and the cross-frequency structure-aware recovery block, the network obtains deep features containing fog degradation correction information and structure enhancement information.

[0237] To gradually restore the spatial resolution of the image, this invention designs a decoder that restores the spatial size step by step through upsampling operations, and achieves multi-scale information fusion by combining shallow features from the encoding stage.

[0238] Specifically, for the bottleneck layer output features B represents the batch size, C represents the number of channels, H represents the height, and W represents the width. First, the feature space size is restored through an upsampling operation: ,

[0239] in, This indicates an upsampling operation, used to improve the spatial resolution of the feature map. The representative uses operational recovery features. This represents the output characteristics of the bottleneck layer.

[0240] Subsequently, the upsampled features are fused with the shallow features at the corresponding scale from the encoding stage:

[0241] ,

[0242] in: This indicates the skip connection features in the encoder output; Concat is used for concatenation along the channel dimension. This represents a 1×1 convolution.

[0243] The fused features are then input into the CFSARB module: ,in This represents the decoded recovery features obtained after feature fusion and cross-frequency structure-aware recovery block processing; This represents the feature obtained by fusing upsampled features with skip connection features from the encoding stage; This indicates a cross-frequency structure sensing recovery block, used for further frequency domain recovery and spatial structure compensation.

[0244] Through the above process, the decoder restores the feature space resolution step by step, and uses the structural information retained in the encoding stage to compensate for the loss of detail that may occur during the upsampling process. After two upsampling operations, the space size is restored to H×W.

[0245] Finally, a 3×3 convolutional layer is used to map the decoded features restored to the input size to the final three-channel dehazing result:

[0246] ,

[0247] in, This represents the final output three-channel dehazed image; This represents a 3×3 convolution operation used to recover features from decoding. Mapped to RGB image space.

[0248] This decoding process can effectively integrate low-frequency color recovery information with high-frequency texture structure information, improving image clarity and detail retention in complex foggy environments.

[0249] Step 5: Train and experimentally validate the image dehazing network, and evaluate the dehazing effect.

[0250] To verify the effectiveness of the proposed image dehazing method based on cross-frequency structure-aware restoration network in complex fog environments, the image dehazing network constructed in steps 1 to 4 was applied to a publicly available image dehazing dataset for training and testing. Objective evaluation metrics and visualized restoration results were used to analyze the dehazing effect.

[0251] Specifically, in terms of data acquisition and processing, this embodiment uses the Haze-4k image dehazing dataset as the experimental dataset. This dataset contains foggy images and their corresponding clear images, which can be used to supervise the training of the image dehazing network. Before the experiment, foggy images and their corresponding clear images in the dataset were read in pairs, ensuring a one-to-one correspondence between the foggy and clear images in terms of spatial size and pixel location. Subsequently, the image data was normalized to map the image pixel values ​​to a uniform numerical range. During the training phase, data augmentation operations such as random cropping and random flipping were performed on the input images to improve the network's adaptability to different fog distributions, different texture structures, and different scene content. During the testing phase, no random augmentation was performed; only the test images were input into the trained network for dehazing and restoration.

[0252] Regarding hardware and software configuration, this embodiment uses the experimental environment shown in Table 1 to complete model training, testing, metric calculation, and result saving:

[0253] Table 1 Experimental Environment

[0254]

[0255] In this system, the GPU is used to accelerate the network's forward propagation, backward propagation, and parameter updates; CUDA provides support for GPU parallel computing; and Python and PyTorch are used to implement network structure building, data reading, loss function calculation, model training, and testing and evaluation. It should be noted that the above hardware and software configuration is only one implementation method in this embodiment. In practical applications, it can be adaptively adjusted according to the model size, dataset size, and hardware conditions, and does not constitute a limitation on the scope of protection of this invention.

[0256] For training parameter settings, the foggy image is used as the network input, and the corresponding clear image is used as the supervision label to train the image dehazing network end-to-end. During training, the Adam or AdamW optimizer is used to update the network parameters, and the initial learning rate can be set to 1×10. -4 The batch size is set according to the GPU memory size, and the input image cropping size can be set to 256×256 or 512×512. During training, pixel-level reconstruction loss is used to constrain the network output, making the dehazed image close to the corresponding sharp image in terms of brightness, color, and structural details. After training, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used as quantitative evaluation metrics for the dehazing effect. PSNR measures the pixel reconstruction error between the network's output dehazed image and the corresponding sharp image; a higher value indicates a smaller recovery error. SSIM measures the structural similarity between the dehazed image and the sharp image; a higher value indicates better preservation of image structure.

[0257] The defogging effect is also analyzed by combining the visual recovery results.

[0258] The ablation experiment results are shown in Table 2. In Table 2, the basic U-shaped dehazing network represents the basic encoder-decoder network without the improved module of this paper; Group A represents the basic network with wavelet dual-domain decomposition added; Group B represents the low-frequency subband Q / K / V relationship modeling added on the basis of Group A; Group C represents the HL / LH high-frequency direction enhancement branch added on the basis of Group B; Group D represents the high-frequency structure guidance mechanism added on the basis of Group C; and the complete dehazing network represents the physical fog perception correction module PHAAM further added on the basis of Group D.

[0259] Table 2 Ablation experimental results under different module combinations

[0260]

[0261] As shown in Table 2, the basic U-shaped dehazing network achieves a PSNR of 23.3879 dB and an SSIM of 0.9447 on the Haze-4k test set, which is sufficient for basic dehazing recovery. However, this network mainly relies on spatial domain convolution to extract features, and lacks specific processing for issues such as frequency degradation, edge blurring, and loss of texture details caused by haze. Therefore, there is still room for improvement in the recovery performance.

[0262] To further verify the effectiveness of each component module in this invention, multiple ablation experiments were conducted to analyze each of the invention's key features, including wavelet dual-domain decomposition, low-frequency relationship modeling, high-frequency direction enhancement, high-frequency structure guidance, and physical fog perception correction. All experiments were performed using the same training data, training strategy, and evaluation metrics to ensure the comparability of the results.

[0263] After adding wavelet dual-domain decomposition to the basic network, group A achieved a PSNR of 22.9800 dB and an SSIM of 0.9413. This result indicates that introducing wavelet decomposition alone does not necessarily improve dehazing performance. This is because wavelet decomposition only divides features into low-frequency and high-frequency sub-bands. The low-frequency sub-band mainly carries brightness, color, and haze degradation information, while the high-frequency sub-band mainly carries edge, texture, and local detail information. If only frequency decomposition is performed without designing corresponding recovery mechanisms for the degradation characteristics of different sub-bands, the network merely changes the feature representation and does not effectively solve the problems of low-frequency haze residue and high-frequency structure weakening. Therefore, this experiment demonstrates that wavelet decomposition itself only provides the foundation for frequency domain processing; it must be combined with low-frequency haze degradation recovery and high-frequency structure enhancement mechanisms to fully realize its potential.

[0264] Group B, building upon wavelet domain decomposition, further incorporates low-frequency sub-band Q / K / V relationship modeling, resulting in a PSNR of 23.4332 dB and an SSIM of 0.9443. Compared to Group A, this result demonstrates the positive effect of low-frequency relationship modeling on dehazing restoration. The theoretical basis lies in the fact that haze degradation primarily manifests as overall image brightness attenuation, color shift, and reduced contrast, information mainly concentrated in the low-frequency sub-band. By constructing Q, K, and V relationships in the low-frequency sub-band, the network's ability to model haze residue, brightness variations, and subject structure information in the low-frequency region can be enhanced. This allows the restoration process to no longer rely solely on local convolutional responses but can adaptively adjust the restoration intensity based on the correlation between low-frequency features. Therefore, this module can improve the brightness and color restoration of the main image region. However, since this group has not yet introduced high-frequency directional enhancement and high-frequency structure guidance, the network's restoration of edges, textures, and local details remains insufficient, thus limiting the performance improvement.

[0265] After adding the HL / LH high-frequency directional enhancement branch to group C based on group B, the PSNR increased to 23.6496 dB, and the SSIM increased to 0.9497. This result demonstrates that the high-frequency directional enhancement branch effectively improves the image's structural detail recovery capability. This is because haze weakens the edge and texture responses in images, especially leading to problems such as blurred directional edges and unclear local textures. The HL and LH subbands respectively focus on representing high-frequency structural information in different directions. This invention employs directional stripe depth convolution and local texture enhancement branches for both, which can more effectively recover the horizontal and vertical edge responses while supplementing local texture details. Therefore, the effective structural information in the high-frequency subbands is enhanced, improving image edge continuity and detail clarity, resulting in a significant improvement in SSIM.

[0266] After further incorporating a high-frequency structure guidance mechanism based on group C, group D achieved a PSNR of 25.2266 dB and an SSIM of 0.9551. This result demonstrates that high-frequency structure information not only improves the detail representation of the high-frequency subbands themselves but also provides effective constraints for the low-frequency restoration process. The theoretical basis is that while the low-frequency subbands carry information about the image's main brightness, color, and haze degradation, their structural boundaries are relatively blurred. If the low-frequency restoration process lacks high-frequency edge and texture constraints, problems such as overly smooth main areas, unclear edge transitions, and weakened structural contours can easily occur. This invention utilizes high-frequency structure guidance maps... By incorporating enhanced high-frequency directional structural information into the low-frequency relation recovery process, the recovery of low-frequency content is constrained by real edge and texture information. Therefore, this module can better preserve image contours, edge continuity, and overall structure while removing haze, which is the main reason for the significant improvements in both PSNR and SSIM.

[0267] The complete model, after adding the Physical Fog Perception Correction (PHAAM) module to group D, achieved a PSNR of 26.8306 dB and an SSIM of 0.9629, the highest among all experimental groups. This result demonstrates that the PHAAM module can further enhance the network's ability to model non-uniform fog degradation. This is because fog concentration, brightness attenuation, and color shift are inconsistent across different image regions; relying solely on ordinary feature learning can easily lead to problems such as localized residual fog, insufficient color recovery, or overcorrection of brightness. PHAAM models the overall fog color shift through a global fog color branch, estimates the degradation degree of different regions through a spatial fog density branch, and further constructs a degradation feature D, giving the low-frequency recovery process a more explicit physical degradation prior. Introducing this degradation feature into the K branch allows low-frequency relationship modeling to focus more on fog residual areas; simultaneously, the high-frequency structure guides the graph... By constraining the V branch, the restoration process balances haze removal and structure preservation. Therefore, the complete model can simultaneously improve low-frequency color brightness recovery, high-frequency structural detail preservation, and overall image sharpness.

[0268] In summary, ablation experiments demonstrate that the modules of this invention are not simply superimposed, but rather form a complementary relationship around different degradation factors in image dehazing. Specifically, wavelet dual-domain decomposition provides the basis for differentiated processing of low and high frequencies; low-frequency relationship modeling enhances the brightness and color recovery of the main haze region; HL / LH high-frequency direction enhancement restores directional edges and local textures; the high-frequency structure guidance mechanism feeds back reliable edge structures to the low-frequency recovery process; and PHAAM further introduces global haze color and spatial haze density priors to improve degradation perception capabilities under non-uniform haze environments. The synergistic effect of these modules improves the network's performance in haze removal, color correction, structure preservation, and detail recovery.

[0269] Figure 7 This is a visualization of the dehazing results of the method of this invention on the Haze-4k test set. From Figure 7 As can be seen, the input foggy image exhibits obvious problems such as fog and haze obstruction, insufficient contrast in the distant view, grayish colors, and blurred edge contours. After processing by the method of this invention, the fog and haze interference in the defogging image is significantly reduced, the colors of the sky, roads, and building areas are more natural, the contrast of the distant view is improved, and structural information such as road edges, building outlines, and target object boundaries are better preserved.

[0270] The visualization results correspond to the module design of this invention. The low-frequency subband recovery branch mainly addresses the problems of brightness attenuation, color shift, and overall contrast reduction caused by haze, thus improving the overall graying of the image and the phenomenon of residual fog occlusion; the HL / LH high-frequency directional enhancement branch mainly addresses the problems of directional edge weakening and texture blurring caused by haze, thus improving the clarity of road edges, building outlines, and target boundaries; the high-frequency structure guidance mechanism introduces the enhanced edge and texture information into the low-frequency recovery process, making the low-frequency brightness and color recovery subject to structural constraints, thereby reducing over-smoothing and outline weakening problems; the physical fog perception correction module improves the network's ability to recover non-uniform haze areas by modeling global fog color and spatial fog density.

[0271] therefore, Figure 7 This study not only demonstrates the dehazing visual effect of the method proposed in this invention, but also subjectively verifies the effectiveness of the inventive points, such as wavelet dual-domain decomposition, low-frequency relationship modeling, high-frequency direction enhancement, high-frequency structure guidance, and physical fog perception correction. The results show that the method of this invention can maintain image edges, textures, and overall structure while reducing fog obstruction, improving the clarity, color realism, and structural integrity of dehazed images. Furthermore, the results demonstrate that the method of this invention can not only improve objective indicators such as PSNR and SSIM, but also enhance image clarity, color realism, and structural integrity in terms of visual effects.

[0272] Example 2:

[0273] This invention provides an electronic device, including a processor and a memory. The memory stores program instructions for implementing the image dehazing method based on physical fog perception and wavelet dual-domain structure guided by Embodiment 1. The processor executes the program instructions stored in the memory to perform image dehazing.

[0274] The processor can also be called a Central Processing Unit (CPU). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0275] The methods described in Embodiment 1 of this invention are applicable to this electronic device embodiment. The specific functions implemented by this electronic device embodiment are the same as those in the above method embodiment, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0276] Example 3:

[0277] This embodiment provides a storage medium storing a computer program or instructions capable of implementing the above-described image dehazing method based on physical fog perception and wavelet dual-domain structure guidance. When the computer program or instructions are run on a computer, the steps of the above-described image dehazing method based on physical fog perception and wavelet dual-domain structure guidance are executed.

[0278] The computer program or instructions can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0279] The methods described in Embodiment 1 of this invention are all applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0280] This invention provides an image dehazing method based on physical fog perception and wavelet dual-domain structure guidance. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. An image dehazing method based on physical fog perception and wavelet dual-domain structure guidance, characterized in that, Includes the following steps: Step 1: Obtain the preprocessed input foggy image; Step 2: Input the foggy image into the encoder, and process it through multi-level feature encoding modules and downsampling operations to obtain encoded features; Step 3: Input the encoded features into the cross-frequency structure-sensing recovery block for processing to obtain the bottleneck layer features; Step 4: Input the bottleneck layer features into the decoder, process them through multi-level upsampling operations, skip connections, and cross-frequency structure-aware recovery blocks, and output a dehazed image.

2. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 1, characterized in that, In step 2, the feature encoding module processes the features, including: First, the input features are processed using 1×1 convolution to obtain the mapped features; Then, depthwise convolution is used to process the mapped features to obtain spatial local features; Subsequently, the local spatial features are input into the SimpleGate gate unit for feature filtering to obtain the filtered features; Finally, the filtered features are processed by 1×1 convolution, and then residual fusion is performed with the input features by introducing a learnable scaling parameter to obtain the output features.

3. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 1, characterized in that, In step 3, the process of processing features in the cross-frequency structure-sensing recovery block includes: First, the input features are normalized and then fed into the wavelet structure-guided recovery module for processing to obtain the wavelet structure-guided recovery features. Then, the wavelet structure-guided recovery features are residually connected with the input features to obtain the first-stage output features; Next, the output features from the first stage are normalized and then input into the scale-adaptive structure selection block for processing to obtain spatial structure compensation features. Finally, the spatial structure compensation features are residually connected with the first-stage output features to obtain the second-stage output features.

4. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 3, characterized in that, The process of feature processing by the wavelet structure-guided recovery module includes: First, the input features are decomposed using discrete wavelet transform to obtain low-frequency sub-band, first high-frequency sub-band, second high-frequency sub-band and third high-frequency sub-band. Secondly, after feature mapping of the low-frequency subband, it is input into the physical fog perception correction module for processing to obtain the degradation features and the low-frequency features after physical fog perception correction. Then, the first high-frequency sub-band and the second high-frequency sub-band are respectively input to the corresponding high-frequency direction enhancement branches for processing to obtain the enhanced first-direction high-frequency features and the second-direction high-frequency features; Next, the high-frequency features in the first and second directions are processed by the orientation calibration structure guidance submodule and then concatenated. After concatenation, they are processed sequentially by 1×1 convolution, depthwise convolution, GELU activation function, 1×1 convolution and Sigmoid function to obtain the structure response map. By introducing 1×1 convolution to perform channel mapping and feature alignment on the structure response map, a high-frequency structure guidance map is obtained. Subsequently, the high-frequency structure guidance map, degradation features, and low-frequency features corrected by physical fog perception are input into the low-frequency relationship enhancement module for processing to obtain the enhanced low-frequency features; The low-frequency relationship enhancement module's feature processing includes: sequentially processing the low-frequency features corrected for physical fog perception through group normalization, 1×1 convolution, and depthwise convolution to obtain normalized low-frequency features; processing the normalized low-frequency features to generate query features and value features; processing the features concatenated with the degenerate features and normalized low-frequency features to obtain key features; performing element-wise operations on the query features and key features after lightweight convolution mapping to obtain low-frequency relationship responses; performing element-wise operations on the high-frequency structure guidance graph and value features to obtain structure-modulated value features; and performing element-wise operations on the structure-modulated value features and low-frequency relationship responses, followed by residual fusion with the low-frequency features corrected for physical fog perception to obtain enhanced low-frequency features. Finally, the enhanced low-frequency features, high-frequency features in the first direction, high-frequency features in the second direction, and high-frequency sub-band in the third direction are spliced ​​together. The spliced ​​features are then processed by inverse discrete wavelet transform to obtain the output features.

5. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 4, characterized in that, The process of the physical fog perception and correction module processing features includes: First, the input features are fed into the global haze branch, and then processed sequentially through global average pooling, 1×1 convolution, GELU activation function, and hyperbolic tangent function mapping to obtain the global haze features: Secondly, the input features are input into the spatial fog density branch and processed sequentially through depthwise convolution, GELU activation function, 1×1 convolution and Sigmoid function to generate spatial fog density response and apply minimum constraint to it. Then, the constrained spatial fog density response is expanded in the channel dimension to obtain spatial fog density features. Then, degradation features are constructed using global fog degradation features and spatial fog density features, and are expressed as follows: , A characteristic of degradation, This is a global fog degradation feature. For element-wise multiplication, Characteristics of spatial fog density; Subsequently, the input features and degradation features are concatenated and then input into the degradation correction branch. The concatenated features are processed sequentially through 1×1 convolution, depthwise convolution, nonlinear activation function and 1×1 convolution to generate correction residuals. Finally, the corrected residuals are fused with the input features using a learnable scaling factor to obtain the output features.

6. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 4, characterized in that, The process of enhancing the branch processing features in the high-frequency direction includes: First, the input features are processed by 1×1 convolution and then fed into a dual-branch structure for processing. One branch generates directional structure features through directional convolution, while the other branch generates local texture features through depth convolution, SimpleGate gate unit and 1×1 convolution in sequence. Then, the directional structural features and local texture features are concatenated and fused using a 1×1 convolution to obtain intermediate features; Next, the intermediate features after gating are multiplied by the learnable scaling factor, and then added to the input features through residual connection to obtain the output features.

7. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 4, characterized in that, The process of the orientation calibration structure guidance submodule processing features includes: The input features are processed sequentially through 1×1 convolution and directional convolution to obtain initial directional structure features; then, the initial directional structure features are processed using depthwise convolution to obtain directional structure features. The directional structural features are input into the channel-gated branch and processed by global average pooling, 1×1 convolution, GELU activation function, 1×1 convolution and Sigmoid function to obtain the channel-gated weights; The directional structural features are input into the sub-band spatial gating branch, and after being processed by channel-dimensional average pooling and channel-dimensional max pooling respectively, feature concatenation, 7×7 convolution and Sigmoid function processing are performed sequentially to obtain the sub-band spatial gating weights. The initial directional structural features are weighted element-wise using channel gating weights and subband spatial gating weights, and then channel fusion is performed using 1×1 convolution to obtain weighted directional structural features. The weighted directional structural features are added to the input features by residual addition to obtain the output directional structural features.

8. The image dehazing method based on physical fog perception and wavelet dual-domain structure guidance according to claim 3, characterized in that, The process of the scale-adaptive structure selection block for feature processing includes: First, multi-scale feature extraction is performed on the input features after 1×1 convolution, and the extracted multi-scale structural features are concatenated to obtain the concatenated features. Then, the concatenated features are processed sequentially through global average pooling, 1×1 convolution, non-linear activation function, 1×1 convolution and Sigmoid function to obtain scale-adaptive weights; Next, the multi-scale structural features are weighted using scale-adaptive weights, and the weighted multi-scale structural features are then processed by feature concatenation, global average pooling, and 1×1 convolution to obtain the compensated features. Finally, the compensated features and input features are residually fused to obtain the output features.

9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the image dehazing method based on physical fog perception and wavelet dual-domain structure as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The system stores a computer program or instructions that, when run on a computer, execute the steps of the image dehazing method based on physical fog perception and wavelet dual-domain structure as described in any one of claims 1 to 8.