A non-uniform fog and haze degraded image restoration method based on semi-supervised learning and related device

Abstract

The application provides a non-uniform haze degraded image restoration method based on semi-supervised learning and a related device, and belongs to the technical field of image processing.The method of the application first uses a channel parallel gate double residual module and a haze perception module of a density perception prior student model to realize pixel-level haze density estimation of a non-uniform haze degraded image and generate a guide image; then the guide image and the original non-uniform haze degraded image are input into a trained diffusion refinement student model, the diffusion refinement student model is constructed based on a multi-head selective scanning mechanism and a conditional diffusion model constructed by a Mamba state space double-branch gate module, a U-shaped noise estimation network is used to iteratively denoise a noisy image, and a restored non-uniform haze degraded image is output. Through the semi-supervised learning framework and the conditional diffusion model, the application solves the problem that paired data sets are difficult to obtain in a real non-uniform haze environment, resulting in poor non-uniform haze degraded image restoration effect.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method and related apparatus for restoring non-uniform haze-degraded images based on semi-supervised learning. Background Technology

[0002] Haze is formed by the interaction of aerosol particles suspended in the near-surface atmosphere with water vapor, leading to severe degradation of image quality captured by image sensors. This degradation mechanism essentially disrupts the stable mapping between low-level pixel representations and high-level semantic meaning, thus weakening the effective transmission of high-frequency information in the scene. Specifically, this manifests as decreased edge sharpness, blurred texture details, distorted color reproduction, and damage to local phase consistency and spectral energy distribution characteristics. These deep-seated image information losses pose a significant technical challenge to computer vision tasks that rely on high-quality visual input. Existing robust restoration models that effectively resist this type of degradation mostly rely on high-quality paired data. However, haze in real-world environments exhibits high randomness and spatiotemporal non-stationarity, making it extremely difficult to obtain accurately paired real-world datasets. Therefore, overcoming the strong dependence on paired data and constructing high-quality image restoration methods suitable for complex real-world scenarios has become a critical problem urgently needing to be solved in this field.

[0003] Currently, existing methods for restoring haze images can be mainly divided into two categories: methods based on physical models and methods based on deep learning.

[0004] Physical model-based methods are primarily built upon the physical foundation of atmospheric scattering models. These models describe the degradation process of haze images as a linear superposition of direct attenuation and atmospheric light scattering terms, achieving image dehazing by estimating prior information such as transmittance maps. However, the performance of these methods is highly dependent on the universality of the pre-existing assumptions and lacks effective modeling capabilities for non-uniform, spatially varying haze distributions. They struggle to accurately describe the complex spatial heterogeneity of haze in real-world scenes, resulting in suboptimal restoration performance in complex natural environments.

[0005] Deep learning-based methods automatically learn the complex nonlinear mapping relationship between hazy and clear images through neural networks, thus achieving image dehazing. Based on different model architectures and learning mechanisms, existing deep dehazing methods can be mainly divided into three categories: methods based on convolutional neural networks (CNNs), methods based on attention mechanisms (Transformer methods), and methods based on generative models.

[0006] Convolutional neural network (CNN)-based methods, through stacked convolutional layers and nonlinear activation functions, can learn complex nonlinear mappings from hazy images to clear images from large-scale data, or accurately estimate transmittance in atmospheric scattering models to achieve image dehazing. However, the local receptive field characteristics of CNN-based methods limit the effective modeling of global image context information, making it difficult to maintain the overall consistency of restoration results when dealing with large-scale non-uniform haze.

[0007] The Transformer method, based on the attention mechanism, calculates the association weights between any pair of pixels in an image through self-attention, directly capturing global contextual information and achieving effective modeling of the spatial distribution of haze. However, when processing high-resolution images, the Transformer architecture's self-attention mechanism introduces quadratic computational complexity, resulting in significant computational overhead. Furthermore, the Transformer is inefficient in capturing local detail features, potentially causing excessive smoothing of local textures in the restored image, affecting visual realism.

[0008] Generative model-based methods aim to learn the data distribution of clear images, thereby generating visually more natural and detailed dehazing results. However, image dehazing methods based on Generative Adversarial Networks (GANs) are unstable during training, susceptible to problems such as mode collapse and gradient vanishing, and the diversity of generated results is often difficult to control.

[0009] The diffusion model learns the data distribution through forward noise addition and backward noise reduction processes, and can restore the detailed structure and texture features of the image with high accuracy step by step. However, its superior performance often depends on the support of large-scale, high-quality paired data, which is difficult to meet in real-world scenarios. Summary of the Invention

[0010] The purpose of this invention is to provide a method and related apparatus for restoring non-uniform haze-degraded images based on semi-supervised learning, which solves the problem that paired datasets are difficult to obtain in real non-uniform haze environments, resulting in poor image restoration effects for non-uniform haze-degraded images.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for restoring non-uniform haze-degraded images based on semi-supervised learning, comprising the following steps: Acquire the original non-uniform haze degradation image; The original non-uniform haze degradation image is input into the trained density-aware prior student model to perform density-aware prior, and a guiding image is obtained. The guide image and the original non-uniform haze degradation image are input into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. Based on the guiding image and the original non-uniform haze degradation image, a U-shaped noise estimation network is used to estimate and denoise the noisy image through iterative calculation, resulting in the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0012] A further improvement of the present invention is that, before inputting the original non-uniform haze degradation image into the density-aware prior student model for density-aware prior and obtaining the guiding image, the original non-uniform haze degradation image is preprocessed. The preprocessing includes cropping, flipping, and rotating geometric transformations on the obtained original non-uniform haze degradation image.

[0013] A further improvement of this invention is that the trained density-aware prior student model is obtained through the following steps: Based on two identical dual-density adaptive augmentation networks, a density-aware prior mean teacher model is constructed, which includes a density-aware prior teacher model and a density-aware prior student model. The density-aware prior student model is trained to obtain a well-trained density-aware prior student model.

[0014] A further improvement of this invention is that the trained diffusion-refined student model is obtained through the following steps: Based on two identical conditional diffusion models, a diffusion-refined mean teacher model is constructed, which includes a diffusion-refined teacher model and a diffusion-refined student model. The diffusion-refined student model is trained to obtain a well-trained diffusion-refined student model.

[0015] A further improvement of this invention is that the density-aware prior student model is trained to obtain the weights of the density-aware prior student model; based on the weights of the density-aware prior student model, the weights of the density-aware prior teacher model are updated using an improved progressive EMA (Exponential Moving Average) strategy to obtain the updated weights of the density-aware prior teacher model. The formula for updating the weights of the density-aware prior teacher model using the improved progressive EMA strategy is as follows:

[0016] in, The weights for the updated density-aware prior teacher model. For the weights of the density-aware prior teacher model, For the weights of the density-aware prior student model, The momentum coefficient; Momentum coefficient The calculation formula is:

[0017] in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This is the maximum value of the momentum coefficient.

[0018] A further improvement of this invention is that the diffusion-refinement student model is trained to obtain the weights of the diffusion-refinement student model; based on the weights of the diffusion-refinement student model, the weights of the diffusion-refinement teacher model are updated using an improved progressive EMA strategy to obtain the updated weights of the diffusion-refinement teacher model. The formula for updating the weights of the diffusion-refined teacher model using the improved progressive EMA strategy is as follows:

[0019] in, To refine the weights of the teacher model in the updated diffusion model, To further refine the weights of the teacher model, To refine the weights of the student model, The momentum coefficient; Momentum coefficient The calculation formula is:

[0020] in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This represents the maximum value of the dynamic momentum coefficient.

[0021] A further improvement of the present invention is that the U-shaped noise estimation network adopts an encoding and decoding structure composed of state-space dual-branch gating blocks.

[0022] Secondly, the present invention provides a non-uniform haze degradation image restoration system based on semi-supervised learning, comprising: The data acquisition module is used to acquire raw, non-uniform haze degradation images; The guided image acquisition module is used to input the original non-uniform haze degradation image into the trained density-aware prior student model to perform density-aware prior and obtain the guided image. The noisy image acquisition module is used to input the guide image and the original non-uniform haze degradation image into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. The restoration module is used to estimate and denoise the noisy image based on the guide image and the original non-uniform haze degradation image by using a U-shaped noise estimation network through iterative calculation, so as to obtain the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0023] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the non-uniform haze degradation image restoration method based on semi-supervised learning described above.

[0024] Fourthly, the present invention provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the non-uniform haze degradation image restoration method based on semi-supervised learning described above.

[0025] Compared with the prior art, the present invention has the following beneficial effects: The proposed method for restoring non-uniform haze degradation images based on semi-supervised learning involves inputting the original non-uniform haze degradation image into a trained density-aware prior student model to obtain a guiding image. This operation can provide reliable pseudo-labels and density-aware prior guidance information for subsequent stages. On the other hand, the guide image and the original non-uniform haze degradation image are input into a trained diffusion-refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, a trained U-shaped noise estimation network is used to estimate the noise randomly sampled from a standard normal distribution. The estimated noise is then removed from the randomly sampled noise, resulting in a noisy image. Based on the guide image and the original non-uniform haze degradation image, the U-shaped noise estimation network iteratively estimates and denoises the noisy image, yielding the restored non-uniform haze degradation image. This operation effectively suppresses the accumulation of confirmation bias and improves the computational efficiency and convergence stability of the diffusion-refinement student model. This effectively solves the problem in existing technologies where paired datasets are difficult to obtain in real non-uniform haze environments, leading to poor restoration results for non-uniform haze degradation images. Furthermore, this invention can achieve a progressive restoration process from coarse to fine through the collaboration of a density-aware prior student model and a diffusion-refinement student model. Meanwhile, this invention constructs a dual-density adaptive enhancement network through a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. Through multi-level feature fusion and attention mechanisms, it achieves adaptive perception and processing of haze concentration differences in different regions of the original non-uniform haze degradation image. By perceiving the haze density distribution and fusing multi-scale features, it generates preliminary restoration results, providing reliable structural prior conditions for subsequent fine processing.

[0026] Furthermore, this invention discloses a U-shaped noise estimation network with an encoding and decoding structure composed of state-space dual-branch gating blocks. This structure, through the synergistic effect of long-range dependency modeling and local feature enhancement, can simultaneously capture global contextual information and local detail features of the image, effectively handle multi-scale degradation patterns in non-uniform haze images, and thus achieve more accurate noise estimation during the diffusion process. Attached Figure Description

[0027] Figure 1 This is a flowchart of the non-uniform haze degradation image restoration method based on semi-supervised learning according to the present invention; Figure 2 This is a schematic diagram of the non-uniform haze degradation image restoration system based on semi-supervised learning according to the present invention; Figure 3 This is a schematic diagram of the non-uniform haze degradation image restoration method based on semi-supervised learning in Embodiment 3 of the present invention; Figure 4This is a network structure diagram of the dual-density adaptive enhancement network DDAE-Net in Embodiment 3 of the present invention; Figure 5 This is a network structure diagram of the U-shaped noise estimation network and the Mamba state space dual-branch gating module in Embodiment 3 of the present invention; Figure 6 This is a comparison of the qualitative evaluation results of the present invention with mainstream image restoration methods on the NH-HAZE dataset, LM-Haze dataset, and RW2AH dataset in Embodiment 3 of the present invention; Figure 7 This is a comparison of the quantitative evaluation results of the method with mainstream image restoration methods on the NH-HAZE dataset in Embodiment 3 of the present invention; Figure 8 This is a comparison of feature point matching results before and after restoring a non-uniform haze-degraded image using the method of the present invention in Example 3; Figure 9 This is a comparison of edge detection results before and after restoring a non-uniform haze-degraded image using the method of the present invention in Example 3; Figure 10 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0028] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0029] Example 1: The flowchart of the non-uniform haze degradation image restoration method based on semi-supervised learning of this invention is as follows: Figure 1 As shown, the non-uniform haze degradation image restoration method based on semi-supervised learning of the present invention includes the following steps: S1. Obtain the original non-uniform haze degradation image; S2. Input the original non-uniform haze degradation image into the trained density-aware prior student model to perform density-aware prior and obtain the guiding image; S3. Input the guide image and the original non-uniform haze degradation image into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, use the trained U-shaped noise estimation network to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. Remove the estimated noise from the noise randomly sampled from the standard normal distribution to obtain the noisy image. S4. Based on the guiding image and the original non-uniform haze degradation image, a U-shaped noise estimation network is used to estimate and denoise the noisy image through iterative calculation to obtain the restored non-uniform haze degradation image; The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0030] Example 2: A schematic diagram of the non-uniform haze degradation image restoration system based on semi-supervised learning of this invention is shown below. Figure 2 As shown, the non-uniform haze degradation image restoration system based on semi-supervised learning of the present invention includes: The data acquisition module is used to acquire raw, non-uniform haze degradation images; The guided image acquisition module is used to input the original non-uniform haze degradation image into the trained density-aware prior student model to perform density-aware prior and obtain the guided image. The noisy image acquisition module is used to input the guide image and the original non-uniform haze degradation image into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. The restoration module is used to estimate and denoise the noisy image based on the guide image and the original non-uniform haze degradation image by using a U-shaped noise estimation network through iterative calculation, so as to obtain the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0031] Example 3: The flowchart of the non-uniform haze degradation image restoration method based on semi-supervised learning of this invention is as follows: Figure 3 As shown, the non-uniform haze degradation image restoration method based on semi-supervised learning of the present invention includes the following steps: S1. Obtain the original non-uniform haze degradation image.

[0032] S2. Input the original non-uniform haze degradation image into the trained density-aware prior student model to perform density-aware prior and obtain the guiding image.

[0033] The process of obtaining the guide image in this step is as follows: A haze density estimation module and a haze perception module are used to generate haze density estimation information and pixel-level haze density perception map, respectively. Based on the haze density estimation information and the original non-uniform haze degradation image, a preprocessed degradation image is obtained. A parallel gated dual residual module with several channels is used to extract features from the preprocessed degradation image to obtain the non-uniform haze degradation image features. The features of the non-uniform haze degradation image and the pixel-level haze density perception map are input into the adaptive quality guidance module to generate a pixel-level confidence map. The generated pixel-level confidence map is used to perform adaptive quality guidance on the original non-uniform haze degradation image and the features of the non-uniform haze degradation image to obtain the perception prior result, which is then used as the guidance image.

[0034] Before obtaining the guiding image, the original non-uniform haze degradation image is preprocessed by inputting it into the density-aware prior student model. The preprocessing includes cropping, flipping, and rotating the obtained original non-uniform haze degradation image.

[0035] Step 2 will be explained in detail below: The density-aware prior student model in this step specifically includes: Based on the Channel Parallel Gate (CP-Gate) module and the Lightweight Gated FeedForward Network (LG-FFN), a Channel Parallel Gate Dual Residual Block (CPGDRB) is constructed. The Channel Parallel Gate module is used for adaptive feature selection, and the Lightweight Gated FeedForward Network is used for cross-channel information interaction. A dual-density adaptive enhancement network (DDAE-Net) is constructed based on a channel-parallel gated dual residual module, a haze density estimator (HDE) module, a haze awareness module (HAM) module, and an adaptive quality-guided module (AQG). The haze awareness module is used to generate a pixel-level haze density awareness map, and the adaptive quality-guided module is used to generate a pixel-level confidence map.

[0036] A density-aware prior student model is constructed based on a dual-density adaptive enhancement network.

[0037] The density-aware prior student model trained in this step is obtained through the following steps: Based on two identical dual-density adaptive augmentation networks, a density-aware prior mean teacher model is constructed, which includes a density-aware prior teacher model and a density-aware prior student model. The density-aware prior student model is trained to obtain a well-trained density-aware prior student model.

[0038] The training of the density-aware prior student model is explained in detail below: The density-aware prior student model is trained to obtain its weights. Based on the weights of the density-aware prior student model, the weights of the density-aware prior teacher model are updated using an improved progressive EMA strategy to obtain the updated weights of the density-aware prior teacher model.

[0039] The formula for updating the weights of the density-aware prior teacher model using the improved progressive EMA strategy is as follows:

[0040] in, The weights for the updated density-aware prior teacher model. For the weights of the density-aware prior teacher model, For the weights of the density-aware prior student model, The momentum coefficient, .

[0041] Momentum coefficient The calculation formula is:

[0042] in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This is the maximum value of the momentum coefficient.

[0043] Step S2 will be explained in detail below: Original non-uniform haze degradation images include labeled paired haze-clean images. Unpaired haze images For labeled data, haze images The density perception prior results are obtained by processing haze density perception through a student network (also called the density perception prior student model, StudentA) based on the haze density perception prior mean teacher model. This result, on the one hand, serves as a pseudo-clear result in the calculation of supervised loss ( On the other hand, it is fed into the diffusion model as a conditional guide, providing prior information for subsequent detailed restoration. For unlabeled data, haze images... After adding noise perturbation, the image is input into the student network (StudentA), along with the original haze image. The data is directly fed into the teacher network (also called the density-perceived prior teacher model, TeacherA) of the haze density perception prior mean teacher model for haze density perception processing, and the consistency loss of the two outputs is calculated. In addition, the results of unlabeled data perceiving haze density through the Teacher Network (TeacherA) In addition to serving as a conditionally guided input diffusion model, it is also used with the original unlabeled haze image. Pseudo-label pairing of haze data Semi-supervised training for the diffusion phase.

[0044] The network structure of the dual-density adaptive enhancement network of this invention is as follows: Figure 4 As shown below, the Dual-Density Adaptive Enhancement Network will be described in detail: This embodiment uses a degraded image as input (the original non-uniform haze degraded image). Density estimation is performed using a Haze Density Estimator (HDE). Global haze features of the image are extracted through multi-layer convolution and pooling operations, and a density parameter is finally output. Where 0 represents no fog and 1 represents dense fog. Subsequently, the density parameter and adaptive weight parameters... Multiplication is performed to scale the input, and density-aware adjustment is applied to the original input. The adjustment process is as follows:

[0045] Adjusted by smog density sensing , through a Convolution maps an image from pixel space to feature space. The feature is then input into a four-layer deep U-shaped encoder-decoder structure consisting of multiple sets of channel-parallel gated double residual blocks (CPGDRB) for feature extraction. The number of channel-parallel gated double residual blocks in each layer of the encoder-decoder structure can be adjusted. During the encoding process, after each encoder layer encodes the input feature, the spatial size of the feature map is downsampled to half of its original size, while the number of channels is doubled. Similarly, each layer of the decoder decodes the input features, upsampling the spatial size to twice its original size, while reducing the number of channels to half its original size.

[0046] The following is a detailed description of the channel parallel gated dual residual module: The Channel Parallel Gate Dual Residual Block is a core component of the encoding and decoding layer. Input features are first normalized, then expanded in channel dimension by a 1×1 convolutional layer, followed by spatial feature extraction using a 3×3 depthwise separable convolution. The extracted features are then fed into the Channel Parallel Gate to achieve adaptive feature selection.

[0047] The channel-parallel gated dual residual module adopts a dual-branch structure: the upper branch performs feature transformation using the LeakyReLU activation function, and the lower branch generates gate weights using the Sigmoid function. The gated features are then weighted by a channel attention mechanism, compressed back to the original feature dimension via 1×1 convolution, and dynamically connected to the input features using residuals. Finally, a lightweight gated feedforward network (LG-FFN) is used to expand the channels of the normalized features, which are then compressed back to the original number of channels after channel-parallel gated processing, promoting cross-channel information interaction.

[0048] After the last layer of feature decoding in the decoder, use a The convolution maps the high-level features processed by the encoder-decoder back to the pixel space, and combines them with the initial features that have not been encoded or decoded. The processed features are obtained by adding the residuals. Subsequently, the haze density sensing module processed the original input. Perform haze sensing and generate pixel-level haze density maps. The density map and the features processed by the encoder and decoder After concatenation along the channel dimension, the data is input into an Adaptive Quality Guiding (AQG) module for processing to obtain a pixel-level confidence map. A value close to 1 indicates high reliability of the encoding / decoding network processing results. Ultimately, based on the haze density map and pixel-level reliability map... and Adaptive weighted fusion is performed to achieve regional differentiation processing and obtain more accurate prior information for haze density perception (also called guiding image). The specific integration process is as follows:

[0049] S3. Input the guide image and the original non-uniform haze degradation image into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, use the trained U-shaped noise estimation network to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. Remove the estimated noise from the noise randomly sampled from the standard normal distribution to obtain the noisy image.

[0050] This step involves diffusing and refining the student model, specifically including: Based on the multi-head selective scanning mechanism, a multi-head selective scanning module (MH-SSM) is constructed. Based on the multi-head selective scanning module, a Mamba State-space Dual-Branch Gate Block (MSDGB) is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0051] The trained diffusion-refined student model in this step is obtained through the following steps: Based on two identical conditional diffusion models, a diffusion-refined mean teacher model is constructed, which includes a diffusion-refined teacher model and a diffusion-refined student model. The diffusion-refined student model is trained to obtain a well-trained diffusion-refined student model.

[0052] The training of the diffusion-refined student model is explained in detail below: The student model is trained to obtain its weights. Based on the student model weights, the weights of the teacher model are updated using an improved progressive EMA strategy to obtain the updated teacher model weights.

[0053] The formula for updating the weights of the diffusion-refined teacher model using the improved progressive EMA strategy is as follows:

[0054] in, To refine the weights of the teacher model in the updated diffusion model, To further refine the weights of the teacher model, To refine the weights of the student model, The momentum coefficient, .

[0055] Momentum coefficient The calculation formula is:

[0056] in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This is the maximum value of the momentum coefficient.

[0057] In this step, the U-shaped noise estimation network adopts an encoding and decoding structure composed of state-space dual-branch gating blocks. The U-shaped noise estimation network is constructed through the Mamba state-space dual-branch gating module.

[0058] Step S3 will be explained in detail below: Step S3 is similar to the haze density perception stage (step S2). Labeled data is used as a guide condition by the prior haze density perception of the haze density perception student model. The data is then input into the student network of the diffusion-refined mean teacher model (also called the diffusion-refined student model, StudentB) for reconstruction. The output of the student network is compared with the real clear image to calculate the supervised loss. For training labeled data, the prior haze density perception results of the student model are used as guiding conditions. For unlabeled data, a more stable haze density perception result is first obtained using the teacher model to reduce confirmation bias in semi-supervised training. Based on this, the pseudo-label pairs formed by the unlabeled data and the haze density perception stage results are noise-perturbed and then input into the student model (StudentB) for restoration; simultaneously, the corresponding data without perturbation is processed by the teacher model (also called the diffusion refinement teacher model, TeacherB), and the outputs of both are constrained by consistency loss. The parameter update strategy of the mean teacher model in the diffusion refinement stage is the same as that of the haze density perception mean teacher model in step S2.

[0059] The loss function used in this embodiment will be described in detail below: In this embodiment, the loss function is composed of both the haze density perception mean teacher model loss function and the diffusion refinement mean teacher model loss function. The loss function is designed as a multi-component, multi-objective composite function, and the overall loss function of the entire framework is... Loss due to supervision and consistency loss The system consists of two phases: the supervision loss and the consistency loss, which are respectively comprised of the haze density sensing phase and the diffusion refinement phase. The formula for calculating the loss function is as follows:

[0060] in, This represents the consistency loss scaling factor. The loss factor represents the haze density perception loss ratio. The haze density perception stage is performed in pixel space, and the loss function adopts a hybrid calculation method of L1 loss and MSE (Mean Squared Error) loss. The diffusion refinement stage achieves image restoration based on the diffusion model, and the model directly outputs Gaussian noise. Therefore, the loss function of this stage is composed of noise prediction loss (including L1 and MSE loss) and pixel-level loss obtained through reparameterization inversion. The comprehensive loss designed through the above multi-level and multi-stage optimization strategy ensures the stable convergence of the model training process and effectively copes with the complex degradation mode of non-uniform haze images.

[0061] The network structure of the U-shaped noise estimation network and the Mamba state-space dual-branch gating module is as follows: Figure 5 As shown below, the U-shaped noise estimation network and the Mamba state-space dual-branch gating module will be described in detail: The input to the U-shaped noise estimation network consists of three parts: time step t In time step t The noisy latent variables and prior information based on haze density perception (also called guide images) are used. The noisy latent variables and haze density perception prior information are first concatenated along the channel dimension, then mapped to the corresponding feature vector space through an embedding module. This feature is then input into an encoding / decoding structure consisting of Mamba State-space Dual-branch Gate Blocks (MSDGBs) for feature extraction. Each Mamba State-space Dual-branch Gate Block (MSDGB) is subjected to the encoded time step... t Modulation enables precise control over the noise intensity estimated by the diffusion model.

[0062] MSDGB is embedded in time steps Embedding and Feature Embedding X Embedding is used as a dual input. Embedding is encoded into three sets of affine parameters and adaptive gating parameters using an MLP for subsequent conditional modulation and dynamic gating operations. The features, after LayerNorm normalization, are first dynamically adjusted using the affine parameters scale1 and shift1 via Modulate, and then a... After convolutional expansion of channels, use The depthwise separable convolution extracts features, which are then fed into the CP-Gate for feature selection. The processed features are further processed into spatial and channel branches. The spatial branch consists of a Multi-Head Selective Scanning Module (MH-SSM) and a standard feedforward network. MH-SSM groups the input features by channel and dimension, selectively scanning along multiple directions such as horizontal, vertical, and diagonal to achieve wide-area neighborhood information fusion. Subsequently, the standard feedforward network FFN (FFN) enables inter-channel feature interaction and fusion. Before and after the MH-SSM operation, Modulate and time-stepping are performed respectively. t The affine parameters and adaptive gating parameters are dynamically adjusted. Channel branching employs a standard channel attention mechanism to recalibrate inter-channel features. Features from spatial and channel branches are then concatenated along the channel dimension and then... The number of channels in the convolution is adjusted, and a residual connection is made with the input features. To further enhance feature representation, a lightweight gated feedforward network (LG-FFN) is used for nonlinear transformation after the residual connection to complete the final feature optimization. The feature map, after multi-layer encoding and decoding, is then processed... The convolution is restored to the image space and combined with noisy latent variables. The noise added during the diffusion process is obtained after performing residual analysis.

[0063] S4. Based on the guiding image and the original non-uniform haze degradation image, a U-shaped noise estimation network is used to estimate and denoise the noisy image through iterative calculation to obtain the restored non-uniform haze degradation image.

[0064] To verify the effectiveness of the proposed semi-supervised learning-based non-uniform haze image restoration method, this embodiment uses three real-world non-uniform haze image datasets for training and validation. The NTIRE image dehazing challenge datasets from 2018 / 2019 / 2020 / 2023 were combined, and duplicate samples were removed to construct a new NH-HAZE dataset. This dataset contains 170 pairs of non-uniform haze-clear image pairs, with resolutions covering 6000×4000, 1600×1200, and other high-resolution specifications. This embodiment uses 157 pairs as the training set and the remaining 13 pairs as the test set. The LMHaze dataset is a large-scale, high-quality real-world non-uniform haze dataset, containing 3925 training pairs and 1115 test pairs, with a resolution of 1200. The RW2AH dataset contains 1758 pairs of real-world haze images. After deduplication, 780 pairs of images were used as the training set and 90 pairs of images were used as the test set.

[0065] This embodiment uses two NVIDIA GeForce RTX 4090 24GB GPUs and is implemented based on the PyTorch 2.4 framework. During training, a 256×256 image patch is randomly selected from the image for each iteration. Inference is performed using a sliding window approach, with a window size of 256×256 and a sliding step size of 32. After training the haze density perception mean teacher model for 100k iterations, a joint training of the diffusion refinement mean teacher model and the haze density perception mean teacher model is conducted. The DDAE-Net network learning rate is... Using the AdamW optimizer, where the first-order momentum decay coefficient... Second-order momentum decay coefficient The weight decay coefficient is The conditional diffusion model was trained for 200k iterations with a learning rate of... Using the AdamW optimizer, where the first-order momentum decay coefficient... Second-order momentum decay coefficient No weight decay is performed. The batch size is set to 16 during training.

[0066] This embodiment restores the test images from the three datasets mentioned above and performs qualitative and quantitative comparative analysis with supervised, semi-supervised, and unsupervised image dehazing methods (GridDehaze, MSBDN, FAA-Net, 4Kdehazing, USID-Net, Dehamer, DehazeFormer, DiffUIR, MixDehaze, OKNet, DEA-Net, HazeGen, and EENet).

[0067] The qualitative evaluation results of this invention on three different datasets compared with mainstream image restoration methods are as follows: Figure 6 As shown, from Figure 6 It can be seen that, compared with other methods, the non-uniform haze degradation image restored by this invention ( Figure 6 The images marked "Ours" have clear and sharp lines and rich details. They can accurately reconstruct realistic color tones and maintain the uniformity and naturalness of colors within the area. At the same time, the color of the white walls in the scene is also accurately reproduced without any yellow or blue tint. With its ability to accurately estimate and differentiate the smog concentration in different areas, it can effectively restore the outline information of distant building groups while clearly preserving the texture details of nearby vegetation, so that the spatial hierarchy and depth of the overall scene are naturally presented.

[0068] To further quantitatively evaluate the quality of the restoration of non-uniform haze-degraded images using the method of this invention, this embodiment performs statistical averaging analysis on 13 paired non-uniform haze-degraded images. The evaluation is conducted using both a reference evaluation metric based on Peak Signal-to-Noise Ratio (PSNR) and a no-reference evaluation metric based on the Natural Image Quality Evaluator (NIQE). PSNR is obtained by calculating the mean square error between the restored image (the restored non-uniform haze-degraded image) and the ground truth image. A higher PSNR value indicates less distortion and better image quality. NIQE evaluates image quality without requiring comparison with the ground truth image; a lower NIQE value indicates higher overall naturalness of the image (the restored non-uniform haze-degraded image).

[0069] The quantitative evaluation and comparison results of this invention with mainstream image restoration methods on the NH-Haze dataset are as follows: Figure 7 As shown, from Figure 7 It can be seen that the PSNR index of the method of the present invention is the highest, indicating that the method of the present invention can accurately restore the image edge and texture information while removing non-uniform haze, effectively reducing pixel error; the NIQE index of the method of the present invention is the lowest, further indicating that the restoration result of the method of the present invention is more in line with the statistical law of natural scene, and is closest to high-quality natural image in terms of visual naturalness and human subjective visual perception quality, effectively avoiding unrealistic artifacts and distortion.

[0070] In this embodiment, in order to objectively evaluate the impact of the method of the present invention on the feature extraction capability of non-uniform haze degraded images before and after restoration, the scale-invariant feature transform (SIFT) algorithm is used for feature point detection and matching analysis. Figure 8 The comparison of feature matching results before and after restoration is shown. The value marked in the lower left corner represents the number of feature points successfully matched between the corresponding image and the reference image by the SIFT algorithm. This value can effectively reflect the richness and stability of identifiable local features in the image. The higher the value, the stronger the robustness and accuracy of feature matching. Figure 8 (a) shows the feature matching between the non-uniform haze degradation image and the reference image. Figure 8 (b) shows the matching result between the image restored by the method of this invention and the same reference image. From Figure 8 As can be seen, after restoration using the method of this invention, the number of correctly matched feature points by the SIFT algorithm increases significantly. This result demonstrates that the restoration process effectively enhances the structural information and local contrast of the image, improves the detectability and discriminability of feature points, and further verifies the effectiveness of the invention method in restoring image details and improving visual quality.

[0071] To evaluate the effectiveness of the method of this invention in preserving and enhancing edge information after restoring images degraded by non-uniform haze, this embodiment introduces the Canny edge detector to perform edge extraction and comparative analysis on the images before and after restoration. The results are as follows: Figure 9 As shown. Among them, Figure 9 (a) shows the original non-uniform haze degradation image and its edge detection results. Figure 9 (b) shows the image restored by the method of this invention and its corresponding edge detection effect. Figure 9As can be seen, after restoration using the method of this invention, the integrity and continuity of the edge structure extracted by the Canny edge detector are significantly improved, and richer detailed edges are effectively detected. This result indicates that the method of this invention can effectively restore and enhance the structural information and local contrast in the image while removing haze interference, thereby improving the saliency of edge features and benefiting subsequent edge-based image analysis and understanding tasks.

[0072] In summary, the proposed semi-supervised learning-based method for restoring non-uniform haze-degraded images requires only a small amount of labeled data to achieve high-quality restoration of such images. This method not only significantly enhances image contrast and clarity, effectively compensating for the loss of detail and texture information due to haze occlusion, but also further improves target scene perception performance in complex haze environments.

[0073] Example 4: Please see Figure 10 As shown, the present invention also provides an electronic device 100 for restoring non-uniform haze degraded images based on semi-supervised learning; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0074] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the non-uniform haze degradation image restoration method based on semi-supervised learning described in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0075] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0076] The memory 101 in the electronic device 100 stores multiple instructions to implement a non-uniform haze degradation image restoration method based on semi-supervised learning, and the processor 102 can execute the multiple instructions to achieve the following: Acquire the original non-uniform haze degradation image; The original non-uniform haze degradation image is input into the trained density-aware prior student model to perform density-aware prior, and a guiding image is obtained. The guide image and the original non-uniform haze degradation image are input into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. Based on the guiding image and the original non-uniform haze degradation image, a U-shaped noise estimation network is used to estimate and denoise the noisy image through iterative calculation, resulting in the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

[0077] Example 5: If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for restoring non-uniform haze-degraded images based on semi-supervised learning, characterized in that, Includes the following steps: Acquire the original non-uniform haze degradation image; The original non-uniform haze degradation image is input into the trained density-aware prior student model to perform density-aware prior, and a guiding image is obtained. The guide image and the original non-uniform haze degradation image are input into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. Based on the guiding image and the original non-uniform haze degradation image, a U-shaped noise estimation network is used to estimate and denoise the noisy image through iterative calculation, resulting in the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

2. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 1, characterized in that, Before obtaining the guiding image, the original non-uniform haze degradation image is preprocessed by inputting it into the density-aware prior student model. The preprocessing includes cropping, flipping, and rotating the obtained original non-uniform haze degradation image.

3. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 1, characterized in that, The trained density-aware prior student model is obtained through the following steps: Based on two identical dual-density adaptive augmentation networks, a density-aware prior mean teacher model is constructed, which includes a density-aware prior teacher model and a density-aware prior student model. The density-aware prior student model is trained to obtain a well-trained density-aware prior student model.

4. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 1, characterized in that, The trained diffusion-refined student model is obtained through the following steps: Based on two identical conditional diffusion models, a diffusion-refined mean teacher model is constructed, which includes a diffusion-refined teacher model and a diffusion-refined student model. The diffusion-refined student model is trained to obtain a well-trained diffusion-refined student model.

5. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 3, characterized in that, The density-aware prior student model is trained to obtain its weights. Based on the weights of the density-aware prior student model, the weights of the density-aware prior teacher model are updated using an improved progressive EMA strategy to obtain the updated weights of the density-aware prior teacher model. The formula for updating the weights of the density-aware prior teacher model using an improved progressive EMA strategy is as follows: in, The weights for the updated density-aware prior teacher model. For the weights of the density-aware prior teacher model, For the weights of the density-aware prior student model, The momentum coefficient; Momentum coefficient The calculation formula is: in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This is the maximum value of the momentum coefficient.

6. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 4, characterized in that, The diffusion-refined student model is trained to obtain its weights; Based on the weights of the diffusion-refined student model, an improved progressive EMA strategy is used to update the weights of the diffusion-refined teacher model, resulting in the updated weights of the diffusion-refined teacher model. The formula for updating the weights of the diffusion-refined teacher model using the improved progressive EMA strategy is as follows: in, To refine the weights of the teacher model in the updated diffusion model, To further refine the weights of the teacher model, To refine the weights of the student model, The momentum coefficient; Momentum coefficient The calculation formula is: in, This represents the current iteration number. This represents the total number of iterations. This is the minimum value of the momentum coefficient. This is the maximum value of the momentum coefficient.

7. The method for restoring non-uniform haze-degraded images based on semi-supervised learning according to claim 1, characterized in that, The U-shaped noise estimation network adopts an encoding and decoding structure composed of state-space dual-branch gating blocks.

8. A system for restoring non-uniform haze-degraded images based on semi-supervised learning, characterized in that, include: The data acquisition module is used to acquire raw, non-uniform haze degradation images; The guided image acquisition module is used to input the original non-uniform haze degradation image into the trained density-aware prior student model to perform density-aware prior and obtain the guided image. The noisy image acquisition module is used to input the guide image and the original non-uniform haze degradation image into the trained diffusion refinement student model. Using the guide image and the original non-uniform haze degradation image as conditions, the trained U-shaped noise estimation network is used to estimate the noise randomly sampled from the standard normal distribution to obtain the estimated noise. The estimated noise is then removed from the noise randomly sampled from the standard normal distribution to obtain the noisy image. The restoration module is used to estimate and denoise the noisy image based on the guide image and the original non-uniform haze degradation image by using a U-shaped noise estimation network through iterative calculation, so as to obtain the restored non-uniform haze degradation image. The density-aware prior student model specifically includes: Based on a channel parallel gating module and a lightweight gating feedforward network, a channel parallel gating dual residual module is constructed. The channel parallel gating module is used for adaptive feature selection, and the lightweight gating feedforward network is used for cross-channel information interaction. A dual-density adaptive enhancement network is constructed based on a channel-parallel gated dual residual module, a haze density estimation module, a haze perception module, and an adaptive quality guidance module. The haze perception module is used to generate a pixel-level haze density perception map, and the adaptive quality guidance module is used to generate a pixel-level confidence map. A density-aware prior student model is constructed based on a dual-density adaptive enhancement network. The diffusion-refined student model specifically includes: A multi-head selective scanning module is constructed based on the multi-head selective scanning mechanism; Based on the multi-head selective scanning module, a dual-branch gating module for the Mamba state space is constructed. A conditional diffusion model is constructed based on the Mamba state-space dual-branch gating module. Based on the conditional diffusion model, a diffusion-refined student model is constructed.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the non-uniform haze degradation image restoration method based on semi-supervised learning as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the non-uniform haze degradation image restoration method based on semi-supervised learning as described in any one of claims 1 to 7.

Images