A method and system for dehazing real-world scene images with uneven fog concentration distribution

Abstract

This invention relates to a method and system for dehazing real-world images with uneven fog distribution. The method inputs the foggy image to be processed into a trained dehazing model. The model backbone is based on the U-Net architecture and includes an encoder and a decoder. The encoder extracts multi-scale feature maps step-by-step through cascaded stages, and sets up a combination of parallel multi-scale convolution and channel attention in at least one stage to enhance the feature extraction capability for regions with different fog concentrations. The decoder gradually recovers the fog-free image, and sets up a dynamic high-frequency filtering module in at least one stage to enhance high-frequency details. Features from each stage of the encoder are passed to the corresponding stage of the decoder via skip connections for fusion. The three modules of hybrid attention, multi-scale parallel convolution, and high-frequency feature extraction are integrated into U-Net: hybrid attention is placed in a shallow layer to enhance useful features, multi-scale parallel convolution extracts multi-scale information, and high-frequency feature extraction combines Laplacian and attention to enhance edge textures, supplemented by a hybrid loss to improve dehazing quality.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to a method and system for dehazing real-world scene images with uneven fog concentration distribution. Background Technology

[0002] Images, as visual data sets rich in information, play a crucial role in human society and computer vision systems. However, with the rapid development of global industrialization and urbanization, smog caused by industrial waste gas and vehicle exhaust has become a global environmental problem. Under severe weather conditions such as smog, particulate matter suspended in the atmosphere will cause severe scattering and absorption of light, resulting in serious degradation of images captured by imaging devices, such as decreased contrast, color distortion, and blurred details.

[0003] This degradation in image quality poses a significant challenge to various advanced computer vision tasks that rely on high-quality visual input. In several key fields such as road traffic, aerospace, military reconnaissance, and intelligent security, haze images directly reduce visibility, severely impacting the accuracy and reliability of tasks such as target detection, recognition, and tracking. This not only leads to substantial economic losses but also threatens public and personal safety. Therefore, image dehazing technology, as a crucial underlying visual preprocessing task, plays an indispensable role in ensuring the performance of subsequent advanced vision applications.

[0004] Traditional dehazing algorithms rely on physical models based on atmospheric scattering models. These models primarily describe two key physical effects: direct attenuation terms and atmospheric light components that cause image whitening. Based on this physical model, the dehazing task is transformed into solving equations: estimating atmospheric light values ​​and transmittance maps from observed hazy images, and then reversing the model to reconstruct a clear, haze-free image. To effectively constrain this problem, researchers have proposed various prior knowledge based on statistical laws, such as the widely used dark channel prior. These priors act as "rules of thumb," providing crucial additional constraints for estimating model parameters. However, the limitations of these methods are also evident. Their dehazing effect is highly dependent on the applicability of the used priors to the current image. In complex scenes or scenarios that do not conform to prior assumptions (such as the presence of large areas of sky or white objects), model parameter estimation is prone to inaccuracy, leading to poor dehazing results, halo effects, or color distortion, thus limiting generalization ability and robustness.

[0005] To overcome the limitations of traditional physical models, current mainstream image dehazing research has fully shifted to a data-driven deep learning paradigm, with its core architecture primarily revolving around two main types of models: Convolutional Neural Networks (CNNs) and Transformers. CNN-based methods, through their hierarchical convolutional operations, can efficiently learn spatial features from shallow to deep layers, such as edges, textures, and even more complex semantic information, from a large number of hazy-clear image pairs. This powerful local feature extraction capability allows them to effectively estimate the parameters of atmospheric scattering models or directly learn end-to-end mappings from hazy to clear images. However, the inherent local receptive field and translation invariance of CNNs also make them insufficient in modeling long-range image dependencies and global contextual information. Furthermore, the distribution of fog is often closely related to the depth and spatial structure of the entire scene, and this deficiency may lead to unevenness or incomplete detail recovery in wide areas (such as the sky) after dehazing.

[0006] To overcome the limitations of CNNs, researchers have recently begun introducing the Transformer architecture into image dehazing. Leveraging its core self-attention mechanism, the Transformer can directly compute the relationships between all pixel blocks in an image, thus achieving accurate modeling of global contextual information. This capability allows for a more accurate assessment of the fog density distribution across the entire scene and better preservation of detail consistency and coherence during restoration. Currently, Transformer-based models exhibit superior performance and are becoming a research frontier in this field. However, directly applying the Transformer to high-resolution data like images incurs significant computational overhead. Therefore, much current research focuses on achieving the same or better processing results with less training data or parameters. For example, knowledge distillation extracts knowledge from a large, trained model into a smaller one; and randomly discarding some input labels during training reduces computational load. Furthermore, exploring hybrid architectures of CNNs and Transformers aims to combine the efficiency of CNNs in local feature extraction with the advantages of Transformers in modeling global dependencies, striving for a better balance between computational cost and dehazing performance, and driving image dehazing technology towards higher accuracy and stronger generalization capabilities.

[0007] Despite significant progress in deep learning-based dehazing methods, several challenges remain. First, most methods are trained on synthetic data, resulting in a domain gap with real-world haze images and degrading performance in practical deployments. Second, existing models typically assume uniform haze distribution, making it difficult to handle spatially variable, non-uniform haze. Third, their ability to recover from extreme haze concentrations remains limited, often resulting in color distortion or loss of detail. Furthermore, balancing model efficiency with recovery quality is a critical issue in practical applications, and the design of lightweight dehazing networks requires further exploration. Future research may need to further combine the interpretability of physical models with the flexibility of data-driven methods to develop robust dehazing systems that can adapt to different scenarios and balance efficiency and performance. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method for dehazing real-world scene images with uneven fog distribution. It can improve image visibility and clarity while preserving the natural visual features and detail levels of the original scene to the maximum extent, and output a dehazed image with a natural visual appearance and clear details, taking into account the characteristics of uneven fog distribution and complex depth changes in real-world scenes.

[0009] To achieve the above objectives, in a first aspect, the present invention provides a method for dehazing real-world scene images with uneven fog concentration distribution, comprising the following steps: The real-world foggy image to be processed is input into the trained image dehazing model to obtain the corresponding dehazed image; The image dehazing model includes a backbone network based on the U-Net architecture, which includes an encoder network and a decoder network. The encoder network includes multiple cascaded encoding stages for progressively extracting multi-scale feature maps of the foggy image; wherein, in at least one encoding stage, a combination structure of parallel multi-scale convolutional modules and channel attention modules is provided to enhance the feature extraction capability for regions with different fog concentrations. The decoder network includes multiple cascaded decoding stages for progressively recovering a fog-free image based on the multi-scale feature map; wherein, in at least one decoding stage, a high-frequency feature extraction module is provided for extracting and enhancing high-frequency information from the feature map to recover image details; The feature maps generated by the encoder network at each level are passed to the corresponding stages of the decoder network via skip connections for feature fusion.

[0010] Furthermore, the combined structure of the multi-scale convolution module and the channel attention module includes a channel attention module and a multi-scale parallel convolution module; in the shallow stage of the encoder network, the channel attention module is combined with the spatial attention module; in the deep stage of the encoder network, the channel attention module is combined with the multi-scale parallel convolution module.

[0011] Furthermore, the processing flow of the multi-scale parallel convolution module includes: Channel expansion is performed on the input feature map; Multi-scale feature extraction is performed on the expanded feature map using multiple parallel convolutional kernels of different sizes to obtain multi-scale features; the sizes of the parallel convolutional kernels include 1×1, 3×3 and 5×5. Channel compression is performed on the extracted multi-scale features; By using a channel shuffling mechanism, different channels are grouped and rearranged to enhance information interaction between features at different scales.

[0012] Furthermore, the dynamic high-frequency filtering module includes: A multi-scale Laplacian Gaussian filter is used to extract high-frequency edge features from the input feature map at multiple scales. The standard deviation parameter of the multi-scale Laplacian Gaussian filter is set as a learnable parameter to adaptively capture edge information in regions with different fog concentrations. The channel attention mechanism introduces weight allocation as a feature preprocessor to filter input features and automatically identify and enhance important feature channels.

[0013] Furthermore, the multi-scale Laplace Gaussian filter includes at least three Laplace Gaussian filters with different initial standard deviations, which are used to extract fine texture features in the fog region, transition features in the medium fog concentration region, and coarse-grained edge features in the dense fog region, respectively.

[0014] Furthermore, the training process for the image dehazing model includes: Obtain training sample pairs consisting of a foggy image and its corresponding clear, fog-free image; The foggy images in the training sample pairs are downsampled to different degrees to obtain foggy image inputs at multiple scales; The foggy images at the multiple scales are fed into the corresponding encoding stages of the encoder network, respectively. The output images at multiple scales are obtained from the corresponding decoding stage of the decoder network, and a multi-output loss function is constructed based on the output images and the clear, fog-free images at the corresponding scales to constrain the training process of the model.

[0015] Furthermore, the image dehazing model is trained using a hybrid loss function, which includes at least a weighted combination of the L1 loss function, the perceptual loss function, and the frequency domain loss function.

[0016] Secondly, the present invention can also provide a dehazing system for real scene images with uneven fog concentration distribution, including an image acquisition module, an image dehazing model module, a jump connection module and an output module; The image acquisition module is used to acquire real-world foggy images of the scene to be processed; The image dehazing model module contains a pre-trained image dehazing model, which is a U-Net-based backbone network including an encoder and a decoder. The encoder is used to extract multi-scale feature maps of the hazy image step by step. The shallow stage of the encoder integrates channel and spatial attention modules, while the deep stage integrates a combination of channel attention modules and multi-scale parallel convolution modules. The decoder is used to gradually recover the hazy image based on the multi-scale feature maps. The deep stage of the decoder integrates a dynamic high-frequency filtering module to enhance image detail recovery. The skip connection module is used to pass the feature maps of each level of the encoder to the corresponding stage of the decoder for fusion; The output module is used to output the dehazed image finally generated by the decoder.

[0017] Thirdly, the present invention also provides a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and the processor can realize the above-mentioned method for dehazing real scene images with uneven fog concentration distribution when executing part or all of the computer executable program.

[0018] Simultaneously, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, it can implement the above-mentioned method for dehazing real scene images with uneven fog concentration distribution.

[0019] Compared with existing technologies, this invention has at least the following advantages: The multi-scale parallel convolution and channel attention combination structure set in the encoder network can simultaneously capture the feature responses of fog concentrations in different ranges. When facing complex scenes where near-field thin fog and distant dense fog intersect, it avoids color distortion caused by over-processing of the thin fog area, enhances the ability to extract transmittance information in the dense fog area, and alleviates common local residues or halo artifacts in the dehazing results. The dynamic high-frequency filtering module introduced in the decoder network actively enhances high-frequency detail information such as edges and textures in the feature map during the upsampling and image resolution restoration process. This ensures that the final dehazed image maintains clear contours and rich surface textures while eliminating fog interference, effectively suppressing the blurring smoothing or edge ringing effects that are prone to occur in ordinary dehazing methods. Multi-scale feature maps output from each stage of the encoder are passed to the corresponding stage of the decoder for fusion via skip connections. This allows deep semantic information to guide the intensity and scope of haze removal, while shallow high-resolution spatial information assists in the accurate reconstruction of image structures. The combination of the two ensures that the dehazing results perform well in terms of global illumination consistency, color fidelity, and object spatial location accuracy. This invention integrates three modules—hybrid attention, multi-scale parallel convolution, and high-frequency feature extraction—into the Unet network. The hybrid attention module is applied to the shallow layers of the encoder and decoder, enabling the model to more intelligently select and strengthen useful features in the early stages of training, emphasizing important channels and suppressing unimportant channels. The multi-scale parallel convolution module completes the multi-scale information extraction of the image. The high-frequency feature extraction module significantly improves the extraction capability of image edge and texture features through the collaborative multi-scale Gaussian Laplacian filter and channel attention mechanism. This invention adopts a hybrid loss function constraint method to further improve the dehazing quality. Attached Figure Description

[0020] Figure 1 This is an overall framework diagram of the present invention; Figure 2 This is a structural diagram of the hybrid attention component in this invention; Figure 3 This is a detailed structural diagram of the multi-scale parallel convolution part in this invention; Figure 4 This is a structural diagram of the high-frequency feature extraction part in this invention; Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method for dehazing real-world images with uneven fog distribution. Based on the Unet model, it innovates on relevant parts to improve overall dehazing accuracy and detail perception, naming it HEDNet. HEDNet is an encoder-decoder structure. The encoder is mainly used to extract features from the input image, while the decoder uses the extracted features to recover the fog-free image. Their functions differ, and their implementation methods also differ. Three modules are inserted into the encoder, and different feature extraction modules are added progressively in each module. The feature extraction module in the shallow network consists of channel attention and spatial attention. As the network depth increases, a multi-scale parallel convolution module is added to the original simple attention. In the decoder module, a high-frequency feature extractor is added in the deeper layers of the network. Skip connections are used to fuse the multi-level outputs of the encoder with the decoder module to better achieve feature aggregation. A multi-input multi-output approach is adopted, where the input is downsampled to different degrees during model training and fed into the encoders at each level of the network. This coarse-to-fine approach can better recognize the multi-scale information of the image. Similarly, multiple outputs are obtained through various levels of decoders, and a progressive approach is used to ensure that the final dehazed image retains clarity and naturalness.

[0022] like Figure 1 As shown, a highly efficient multi-scale parallel convolution module is introduced into the original Unet backbone network. Compared to the original convolution in the network, which only performs simple tensor numerical operations on the input data, the multi-scale parallel convolution module adopts a channel expansion-multi-scale processing-compression process and utilizes a channel shuffling mechanism to effectively improve cross-scale feature interaction capabilities, achieving high-quality reconstruction of image details and textures. Specifically, the channel shuffling mechanism refers to grouping and rearranging different channels after multi-scale convolution processing, allowing information from each channel to flow between groups, thereby breaking the fixed order between channels and enhancing cross-group feature fusion capabilities. This process does not introduce additional parameters; it improves the diversity of feature representation without increasing computational burden simply by index rearrangement. Furthermore, to provide more informative input for subsequent multi-scale convolution modules, weight allocation is introduced as a feature preprocessor to filter input features, automatically identifying and enhancing important feature channels. Figure 1 In this process, the input features first undergo channel recalibration in a hybrid attention module, and then enter a multi-scale parallel convolution module: the features sequentially undergo channel expansion convolution, multi-scale convolution (containing convolution kernels of different sizes), channel compression, and channel shuffling operations, finally outputting an enhanced feature map of the same size as the input. Residual connections are used between the modules to ensure stable gradient flow.

[0023] like Figure 2As shown, to address the issue of incomplete dehazing, this invention introduces a hybrid attention mechanism module. This module, through the synergistic effect of channel attention and spatial attention, enables the network to adaptively focus on important structural and textural regions in foggy images, enhancing the perception and extraction of key features, effectively suppressing noise interference from the background and uniform regions, and reducing erroneous processing during dehazing. Simultaneously, this module guides the network to generate more discriminative feature representations, promoting high-fidelity restoration of image structure and details, thereby improving the visual quality and naturalness of the dehazing results. Combined with... Figure 2 The specific structure and data operation flow of the hybrid attention mechanism module are explained as follows: The input feature map first passes through the channel attention module, where global average pooling and global max pooling are performed on each channel to generate two feature description vectors. These vectors are then fed into a shared multilayer perceptron and added together. After passing through a sigmoid activation function, the channel attention weights are obtained and multiplied with the original feature map to obtain the channel recalibrated features. These features are then input into the spatial attention module, where average pooling and max pooling are performed on each channel dimension. The two results are concatenated and then passed through a 7×7 convolutional layer and a sigmoid activation function to generate a spatial attention map. This map is then multiplied with the original features to output the final enhanced feature map.

[0024] like Figure 3 As shown, this invention specifically designs three convolutional kernels of different sizes to address the multi-scale characteristics of non-uniform fog. The 1×1 small convolutional kernel focuses on extracting subtle transmittance changes in local areas, suitable for feature capture in light fog regions; the 3×3 medium convolutional kernel can acquire fog concentration distribution features over a medium range; and the 5×5 large convolutional kernel is more suitable for processing features in large-scale dense fog regions. Compared to traditional downsampling methods, this multi-scale parallel processing structure better preserves the detailed information of the fog image, avoids losing important transmittance clues during downsampling, and significantly improves the model's adaptability to different fog distributions. Figure 3 The specific structure and data processing flow are as follows: The input feature map is simultaneously fed into three parallel branches, which perform convolution operations using 1×1, 3×3, and 5×5 convolutional kernels respectively. Each convolutional layer is followed by a batch normalization (BN) layer to accelerate convergence and stabilize the training process, and then a ReLU6 activation function is applied to introduce non-linear expressive power. The output feature maps of the three branches are concatenated and fused along the channel dimension, and then passed through a 1×1 convolutional layer for channel fusion and dimensionality reduction, finally outputting a multi-scale enhanced feature map. The entire module maintains consistent spatial resolution between the input and output, ensuring that no detailed information is lost.

[0025] like Figure 4As shown, this invention considers that combining spatial and frequency domain information yields better dehazing results, and designs a high-frequency feature extraction module that integrates frequency and spatial feature extraction. Haze distributed within an image can dilute useful information within the region, especially high-frequency information, thus affecting visibility and human perception. Therefore, the processing focus is placed on the high-frequency components of the image region. This module employs a multi-scale Laplacian Gaussian (LoG) filter for high-frequency feature extraction. The LoG filter, by combining Gaussian blur and the Laplacian operator, can effectively extract abrupt change features.

[0026] The Laplacian operator is a second-order derivative operator that detects areas of maximum grayscale variation in an image. If the image pixel intensity value is... Its corresponding Laplace operator As shown in the following formula:

[0027] In convolution operations, the Laplacian operator can be approximated using a discrete form of its second derivative to achieve a second-order differential estimate of image pixel intensity. However, directly applying the Laplacian operator for differentiation significantly amplifies noise interference in the image, making the result extremely sensitive to noise. Therefore, a Gaussian-based smoothing preprocessing is typically introduced before the differentiation operation to suppress the influence of noise. The mathematical expression of a two-dimensional Gaussian smoothed convolution kernel is shown below:

[0028] First, Gaussian smoothing is used to ignore noise and redundant information in the image. Then, the Laplacian operator is used to detect rapidly changing parts of the image, i.e., high-frequency information. Since the Gaussian and Laplacian kernels are usually much smaller than the image, and the parameters of the LoG filter kernel can be pre-calculated, only one convolution needs to be performed on the image at runtime. Therefore, this method usually requires very few arithmetic operations. The mathematical expression of the LoG filter is as follows:

[0029] The standard deviation of the Gaussian kernel affects the results of high-frequency information detection. When the standard deviation is small, the effect of Gaussian smoothing is small, and more noise and details in the image are preserved. For this reason, the image is greatly affected by noise when the standard deviation is small. When the standard deviation is large, the effect of Gaussian smoothing is more obvious, which can eliminate a lot of unnecessary noise. However, it should not be too large, otherwise it will cause image distortion and loss of some important information.

[0030] In this invention, to address the problem of dehazing non-uniform haze maps, the following steps are taken: It is a learnable parameter. Three different parallel... A filter that captures edges of varying thickness. Initial value. A value of 0.5 extracts fine textures in hazy areas, 1.0 extracts features between hazy and dense fog, and 1.5 captures coarse-grained edges in dense fog areas. For example... Figure 4 As shown, LoG_0.5 LoG_1.0 And LoG_1.5 These represent three different LoG filter scales.

[0031] The effects of this invention can be further illustrated by the following simulation examples.

[0032] In the simulation of this invention, the computer system is Win10, the simulation environment is a deep learning environment built with PyTorch, and the simulation is performed on four datasets: Dense-Haze, NH-Haze, O-Haze, and I-Haze.

[0033] II. Simulation Content and Results Table 1 compares the performance of HEDNet with other typical methods. It can be seen that HEDNet outperforms all other algorithms.

[0034]

[0035] As can be seen, the model's PSNR is improved by 0.36 dB on the Dense-Haze dataset, 0.31 dB on the NH-Haze dataset, 1.10 dB on the O-Haze dataset, and 0.23 dB on the I-Haze dataset compared to the previous best methods. The bolded results in the table represent the best results, and the italicized results represent the second-best results. From the experiments conducted in this invention, it can be concluded that HEDNet outperforms mainstream image dehazing algorithms in terms of dehazing accuracy and visual appeal, and therefore can be applied to the field of image dehazing.

[0036] Based on the method described in this invention, a dehazing system for real-world scene images with uneven fog concentration distribution can also be provided, including an image acquisition module, an image dehazing model module, a skip connection module, and an output module; The image acquisition module is used to acquire real-world foggy images of the scene to be processed; The image dehazing model module contains a pre-trained image dehazing model, which is a U-Net-based backbone network including an encoder and a decoder. The encoder is used to extract multi-scale feature maps of the hazy image step by step. The shallow stage of the encoder integrates channel and spatial attention modules, while the deep stage integrates a combination of channel attention modules and multi-scale parallel convolution modules. The decoder is used to gradually recover the hazy image based on the multi-scale feature maps. The deep stage of the decoder integrates a dynamic high-frequency filtering module to enhance image detail recovery. The skip connection module is used to pass the feature maps of each level of the encoder to the corresponding stage of the decoder for fusion; The output module is used to output the dehazed image finally generated by the decoder.

[0037] On the other hand, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the method for dehazing real scene images with uneven fog concentration distribution as described in the present invention.

[0038] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the method for dehazing real scene images with uneven fog concentration distribution described in the present invention when executing the computer executable program.

[0039] The computer device may be a laptop, tablet, desktop computer, or workstation.

[0040] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf programmable gate array (FPGA).

[0041] The memory described in this invention can be an internal storage unit of a laptop, tablet, desktop computer, or workstation, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0042] Computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0043] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for dehazing realistic scene images with uneven fog concentration distribution, characterized in that, Includes the following steps: The real-world foggy image to be processed is input into the trained image dehazing model to obtain the corresponding dehazed image; The image dehazing model includes a backbone network based on the U-Net architecture, which includes an encoder network and a decoder network. The encoder network includes multiple cascaded encoding stages for progressively extracting multi-scale feature maps of the foggy image; wherein, in at least one encoding stage, a combination structure of parallel multi-scale convolutional modules and channel attention modules is provided to enhance the feature extraction capability for regions with different fog concentrations. The decoder network includes multiple cascaded decoding stages for progressively recovering a fog-free image based on the multi-scale feature map; wherein, in at least one decoding stage, a dynamic high-frequency filtering module is provided for extracting and enhancing high-frequency information from the feature map to recover image details; The feature maps generated by the encoder network at each level are passed to the corresponding stages of the decoder network via skip connections for feature fusion.

2. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 1, characterized in that, The combined structure of the multi-scale convolution module and the channel attention module includes a channel attention module and a multi-scale parallel convolution module; in the shallow stage of the encoder network, the channel attention module is combined with the spatial attention module; in the deep stage of the encoder network, the channel attention module is combined with the multi-scale parallel convolution module.

3. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 2, characterized in that, The processing flow of the multi-scale parallel convolution module includes: Channel expansion is performed on the input feature map; Multi-scale feature extraction is performed on the expanded feature map using multiple parallel convolutional kernels of different sizes to obtain multi-scale features; the sizes of the parallel convolutional kernels include 1×1, 3×3 and 5×5. Channel compression is performed on the extracted multi-scale features; By using a channel shuffling mechanism, different channels are grouped and rearranged to enhance information interaction between features at different scales.

4. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 1, characterized in that, The dynamic high-frequency filtering module includes: A multi-scale Laplacian Gaussian filter is used to extract high-frequency edge features from the input feature map at multiple scales. The standard deviation parameter of the multi-scale Laplacian Gaussian filter is set as a learnable parameter to adaptively capture edge information in regions with different fog concentrations. The channel attention mechanism introduces weight allocation as a feature preprocessor to filter input features and automatically identify and enhance important feature channels.

5. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 4, characterized in that, The multi-scale Laplace Gaussian filter includes at least three Laplace Gaussian filters with different initial standard deviations, which are used to extract fine texture features in the fog region, transition features in the medium fog region, and coarse-grained edge features in the dense fog region, respectively.

6. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 1, characterized in that, The training process for the image dehazing model includes: Obtain training sample pairs consisting of a foggy image and its corresponding clear, fog-free image; The foggy images in the training sample pairs are downsampled to different degrees to obtain foggy image inputs at multiple scales; The foggy images at the multiple scales are fed into the corresponding encoding stages of the encoder network, respectively. The output images at multiple scales are obtained from the corresponding decoding stage of the decoder network, and a multi-output loss function is constructed based on the output images and the clear, fog-free images at the corresponding scales to constrain the training process of the model.

7. The method for dehazing real-scene images with uneven fog concentration distribution according to claim 1 or 6, characterized in that, The image dehazing model is trained using a hybrid loss function, which includes at least a weighted combination of the L1 loss function, the perceptual loss function, and the frequency domain loss function.

8. A system for dehazing real-world scene images with uneven fog concentration distribution, characterized in that, include: Image acquisition module, image dehazing model module, skip connection module, and output module The image acquisition module is used to acquire real-world foggy images of the scene to be processed; The image dehazing model module contains a pre-trained image dehazing model, which is a U-Net-based backbone network including an encoder and a decoder. The encoder is used to extract multi-scale feature maps of the hazy image step by step. The shallow stage of the encoder integrates channel and spatial attention modules, while the deep stage integrates a combination of channel attention modules and multi-scale parallel convolution modules. The decoder is used to gradually recover the hazy image based on the multi-scale feature maps. The deep stage of the decoder integrates a dynamic high-frequency filtering module to enhance image detail recovery. The skip connection module is used to pass the feature maps of each level of the encoder to the corresponding stage of the decoder for fusion; The output module is used to output the dehazed image finally generated by the decoder.

9. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading part or all of the computer-executable program from the memory and executing it, and when the processor executes part or all of the computer-executable program, it can implement the method for dehazing real scene images with uneven fog concentration distribution as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of a method for dehazing real-world images with uneven fog concentration distribution as described in any one of claims 1-6.

Images