An image defogging method and system based on generative distillation and physical perception guidance

Abstract

The application belongs to the technical field of image processing, and provides an image defogging method and system based on generative distillation and physical perception guidance, which comprises the following steps: obtaining an original foggy image; generating a semantic clear reference image and a pre-defogging geometric structure anchor point based on the obtained original foggy image; extracting cross-appearance matching points according to the generated semantic clear reference image and the pre-defogging geometric structure anchor point; establishing a geometric correspondence between images according to the cross-appearance matching points, performing local appearance migration and overlapping fusion, and obtaining an image pseudo-clear label; forming pseudo-paired data based on the original foggy image and the image pseudo-clear label, training a conditional latent diffusion student model by using the pseudo-paired data, performing physical perception based on the original foggy image, and constructing a spatial guidance intensity map; performing spatial modulation on positive and negative conditional noise prediction according to the spatial guidance intensity map, obtaining a defogging image through reverse diffusion sampling, and completing image defogging based on generative distillation and physical perception guidance.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image dehazing method and system based on generative distillation and physical perception guidance. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Image dehazing is a fundamental task in image restoration, aiming to recover clear scene radiation from degraded images affected by atmospheric scattering. Existing image dehazing methods typically rely on prior estimates of transmittance and atmospheric light, such as dark channels, color attenuation, or nonlocal statistics. When the sky, large white objects, complex light sources, or non-uniform haze do not meet the preset statistical assumptions, color distortion, halos, local over-enhancement, or haze residue can easily occur. In addition, because real haze is also affected by non-uniform illumination, anisotropic scattering, particle scale variations, camera response, and various degradation couplings, there is a significant domain difference between simple synthetic degradation and real hazy images. This leads to incomplete dehazing, texture blurring, color shifts, and artifacts in real scenes, especially in heavily hazy scenes.

[0004] Unpaired learning, domain adaptation, codebook prior, and diffusion-generative dehazing methods can be employed, but unpaired distribution constraints struggle to provide clear supervision for the same scene and pixel location, and the recovery capability of conventional diffusion-generative dehazing methods remains limited by the distribution of training data. Other methods construct more realistic degradation data by learning the fogging process before training the diffusion-generative dehazing model; however, these methods fail to directly leverage the open-world semantic knowledge contained in large image editing models to construct clear supervision. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an image dehazing method and system based on generative distillation and physical perception guidance. Starting from a real, original foggy image, it obtains a semantically clear reference image and pre-dehazing geometric anchor points. Further, it extracts cross-appearance matching points to obtain pseudo-clear image labels. Combining a conditional latent diffusion student model, it constructs a spatial guidance intensity map based on the physical perception of the original foggy image and spatially modulates the noise prediction during the diffusion sampling process, thus completing image dehazing based on generative distillation and physical perception guidance. This solves problems such as missing real foggy-fog-free pairing data, geometric misalignment of image editing results, local appearance migration artifacts, high deployment costs of complex teacher pipelines, and the difficulty of uniform diffusion guidance adapting to non-uniform fog.

[0006] According to some embodiments, the first aspect of the present invention provides an image dehazing method based on generative distillation and physical perception guidance, employing the following technical solution: An image dehazing method based on generative distillation and physical perception guidance includes: Obtain the original foggy image; Based on the acquired original foggy image, a semantically clear reference image and pre-dehazed geometric anchor points are generated. Based on the generated semantically clear reference image and pre-dehazed geometric anchor points, extract cross-appearance matching points; Based on the extracted cross-appearance matching points, establish the geometric correspondence between images, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels; Pseudo-paired data is formed based on the original foggy image and the pseudo-clear image labels. The conditional latent diffusion student model is trained using the pseudo-paired data. Physical perception is performed based on the original foggy image to construct a spatial guidance intensity map. Spatial modulation is performed on the positive and negative conditional noise predictions based on the spatial guidance intensity map, and the dehazed image is obtained by back diffusion sampling, thus completing the image dehazing based on generative distillation and physical perception guidance.

[0007] As a further technical limitation, the acquired original foggy image is processed using dehazing editing instructions to generate a semantically clear reference image; the dehazing editing instructions include at least removing the effects of fog and atmospheric scattering to enhance the texture and contrast of the visible area, maintaining the scene category, image aspect ratio, basic semantics of objects and original lighting context, and maintaining the original color context; the generated pre-dehazing geometric anchor points are consistent with the pixel coordinates, object boundaries, relative positions and scene layout of the original foggy image.

[0008] As a further technical limitation, in the process of extracting cross-appearance matching points, both the semantically clear reference image and the pre-dehazed geometric anchor points are input into the cross-appearance image matching model to obtain cross-appearance matching points; the cross-appearance image matching model adopts the MatchAnything model.

[0009] As a further technical limitation, in the process of obtaining the pseudo-sharp image label, overlapping local windows are set on the pre-dehazed geometric anchor points according to a preset window size and step size. The corner points of the local windows are mapped onto the semantically sharp reference image through a displacement field. The corresponding region of the semantically sharp reference image is determined based on the mapped corner points. Local statistical migration is performed across appearance matching points in the determined region. The local statistical migration values ​​are overlapped and fused to obtain a fused aligned image. The obtained fused aligned image is repaired using a truncation diffusion process to obtain the pseudo-sharp image label.

[0010] As a further technical limitation, during the physical perception process of the original foggy image, conditional backdiffusion sampling is performed on the original foggy image to obtain positive and negative conditional noise predictions of the image; the image transmittance is estimated based on the original foggy image to obtain a spatial guidance intensity map; the positive and negative conditional noise predictions are spatially modulated according to the spatial guidance intensity map to obtain noise predictions used to update latent variables.

[0011] As a further technical limitation, pre-dehazing geometric anchor points are generated based on the original foggy image and the structure-preserving dehazing model; wherein, the structure-preserving dehazing model adopts a conditional diffusion-based dehazing model or a geometry-based dehazing model.

[0012] According to some embodiments, a second aspect of the present invention provides an image dehazing system based on generative distillation and physical perception guidance, employing the following technical solution: An image dehazing system based on generative distillation and physical perception guidance includes: The acquisition module is configured to acquire the raw, foggy image. The generation module is configured to generate a semantically clear reference image and pre-dehazed geometric anchor points based on the acquired original hazy image; The extraction module is configured to extract cross-appearance matching points based on the generated semantically clear reference image and pre-dehazed geometric anchor points; The fusion module is configured to establish geometric correspondences between images based on the extracted cross-appearance matching points, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels. The adjustment module is configured to form pseudo-paired data based on the original foggy image and the image pseudo-clear labels, use the pseudo-paired data to train the conditional latent diffusion student model, perform physical perception based on the original foggy image, and construct a spatial guidance intensity map. The dehazing module is configured to spatially modulate the positive and negative conditional noise predictions based on the spatial guidance intensity map, obtain the dehazed image through backdiffusion sampling, and complete the image dehazing based on generative distillation and physical perception guidance.

[0013] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium, employing the following technical solution: A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of an image dehazing method based on generative distillation and physical perception guidance as described in the first aspect of the present invention.

[0014] According to some embodiments, the fourth aspect of the present invention provides an electronic device, which adopts the following technical solution: An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of an image dehazing method based on generative distillation and physical perception guidance as described in the first aspect of the present invention.

[0015] According to some embodiments, the fifth aspect of the present invention provides a computer program product, which adopts the following technical solution: A computer program product includes software code, wherein the program in the software code performs the steps of an image dehazing method based on generative distillation and physical perception guidance as described in the first aspect of the present invention.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention is based on original foggy images, preserving the real atmospheric scattering, complex lighting, and camera imaging characteristics, reducing the "domain gap" caused by synthetic degradation; through a decoupling mechanism where appearance is taken from the generation reference and geometry is taken from the structural anchor point, the open-world semantic prior is converted into structural consistency supervision that can be used for image restoration; through the complex mechanism of the data pipeline, the stability of cross-appearance matching is improved, further mitigating artifacts and color regression caused by block migration; through teacher-student distillation, the multi-stage offline pipeline is compressed into a single deployable model, reducing the complexity of online inference; and the transmittance physical prior is directly embedded into diffusion sampling, so that the spatial intervention of the generation capability matches the reliability of the observation signal, taking into account both the restoration of heavy fog and the fidelity of light fog areas. Attached Figure Description

[0017] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0018] Figure 1 This is a flowchart of an image dehazing method based on generative distillation and physical perception guidance according to Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of an image dehazing method based on generative distillation and physical perception guidance according to Embodiment 1 of the present invention; Figure 3 This is a structural block diagram of an image dehazing system based on generative distillation and physical perception guidance, as shown in Embodiment 2 of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0022] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0023] In this invention, terms such as "fixed connection," "connected," and "linked" should be interpreted broadly, indicating a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can determine the specific meaning of these terms in this invention based on the specific circumstances, and they should not be construed as limitations on the invention.

[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0025] Example 1 Embodiment 1 of this invention introduces an image dehazing method based on generative distillation and physical perception guidance.

[0026] like Figure 1 The image dehazing method shown includes: Obtain the original foggy image; Based on the acquired original foggy image, a semantically clear reference image and pre-dehazed geometric anchor points are generated. Based on the generated semantically clear reference image and pre-dehazed geometric anchor points, extract cross-appearance matching points; Based on the extracted cross-appearance matching points, establish the geometric correspondence between images, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels; Pseudo-paired data is formed based on the original foggy image and the pseudo-clear image labels. The conditional latent diffusion student model is trained using the pseudo-paired data. Physical perception is performed based on the original foggy image to construct a spatial guidance intensity map. Spatial modulation is performed on the positive and negative conditional noise predictions based on the spatial guidance intensity map, and the dehazed image is obtained by back diffusion sampling, thus completing the image dehazing based on generative distillation and physical perception guidance.

[0027] like Figure 2 As shown, this embodiment includes three interconnected stages: geometry-appearance decoupling data construction, conditional latent diffusion student model training, and physical perception space adaptive inference. It directly utilizes the original foggy image to construct high-quality pseudo-paired data, achieving single-image dehazing with stable structure, natural color, rich detail, and adaptability to non-uniform real fog without deploying a large image editing model and a multi-stage data processing pipeline. First, the original foggy image is input into the image editing large model to generate a clear appearance reference image with natural colors, clear textures, and reasonable scene semantics. At the same time, a structure-preserving dehazing model is used to generate a pre-dehazing result that is strictly aligned with the original foggy image, which serves as a geometric anchor point.

[0028] Secondly, high-confidence feature correspondences are extracted between structural anchor points and clear appearance reference images. A dense displacement field is constructed using neighborhood displacement weighted interpolation, and local corresponding regions are determined according to this displacement field. The color and contrast statistics of the clear appearance reference image are statistically transferred to the structural anchor points. After fusing the multiple transfer results of overlapping local windows, block artifacts, color anomalies, and local structural artifacts are eliminated by truncation and re-noise addition and conditional backdiffusion, resulting in pseudo-clear labels that are geometrically consistent with the original hazy image. Subsequently, using the pseudo-paired data composed of the original hazy image and the pseudo-clear labels, a latent diffusion student model with conditional control branches is trained for noise prediction, distilling the semantic reconstruction, appearance optimization, and structure preservation capabilities in the complex teacher pipeline into a single student model.

[0029] Finally, during the inference phase, a transmittance map is estimated from the foggy image to be processed, and the transmittance is mapped to a spatially varying classifier-free guided intensity map. At each reverse diffusion time step, the noise prediction difference between the positive and negative conditions is modulated, so that the low transmittance dense fog region has a stronger generation prior intervention, while the high transmittance region maintains a stronger input structure constraint.

[0030] This embodiment divides real-scene image dehazing into three stages: offline teacher data construction, student model distillation training, and online physical perception inference. In the offline stage, starting with the original hazy image, the clear appearance source and the true geometric source are obtained separately. Then, pseudo-clear labels are constructed through cross-appearance matching, dense displacement interpolation, local appearance transfer, and generative artifact repair. In the training stage, the original hazy image and pseudo-clear labels are used to train a conditional latent diffusion student model. In the inference stage, transmittance maps are used to guide the diffusion of the student model through spatial adaptive modulation. This embodiment forms a technical closed loop around the same student dehazing model. The teacher data construction stage endows the student model with strong semantic reconstruction and natural appearance generation capabilities, while the physical perception inference stage limits the degree of intervention of this capability in different spatial regions.

[0031] This embodiment constructs a pseudo-clear label from the original foggy image based on geometry-appearance decoupling, which has the same geometric structure and a natural clear appearance. It does not require additional acquisition of real fog-free images at the same location, time and lighting conditions. Instead, it assigns the clear appearance generation and the preservation of the real structure to different processing branches, and then recombines the two through explicit geometric correspondence.

[0032] In this embodiment, the input original foggy image is assumed to be: It employs a large generative model with image conditional editing capabilities, based on... Generate semantically clear reference images using dehazing editing commands. Among them, the dehazing editing instructions include at least the following three types of constraints: removing the effects of fog and atmospheric scattering, enhancing the texture and contrast of visible areas; maintaining the scene category, image aspect ratio, basic semantics of objects and original lighting context, restricting unfounded content reconstruction for completely invisible areas; maintaining the original color context, avoiding converting night scenes into day scenes, and distinguishing between fog and snow, water splashes, clouds or white objects.

[0033] This embodiment can use the Qwen-Image-Edit image editing model, setting the guidance strength to 4.0 and the inference steps to 50, to generate... ; It is used to provide clear colors, natural textures, and scene semantics, but because there may be changes in object position, contour deformation, or local content reconstruction, it is not directly used as a supervision label for student models.

[0034] The same original foggy image Input the structure-preserving dehazing model to obtain the pre-dehazing geometric anchor points. It should be noted that the obtained pre-defogging geometric anchor points Compared to the original foggy image Maintaining consistency in pixel coordinates, object boundaries, relative positions, and scene layout, thus serving as geometric anchor points for subsequent appearance migration.

[0035] This embodiment can use a conditional diffusion-based dehazing model for pre-dehazing, or it can use convolutional networks, Transformer networks, or other dehazing models that can preserve the input geometry. Setting a pre-dehazing step has two functions: First, and Both appear as relatively clear images, with less difference in appearance between the original foggy image and the original image. The difference in appearance allows the cross-appearance matching model to obtain more and more evenly distributed matching points. Second, Most of the fog texture has been removed, so that subsequent artifact repair only needs to deal with mild alignment degradation, thereby reducing the number of back diffusion steps and the degree to which clear appearance colors regress to the basic dehazing output.

[0036] This embodiment uses pre-defogging geometric anchor points. With semantically clear reference images Inputting a cross-appearance image matching model yields a set of high-confidence matching points, i.e. ;in, Indicates the number of matching points, Indicates the first The matching points are at the anchor points of the pre-dehazing geometry. coordinates in This indicates the corresponding point in the semantically clear reference image. The coordinates in the diagram.

[0037] This embodiment can use the MatchAnything cross-modal image matching model, and can further improve matching reliability by using matching confidence thresholds, bidirectional consistency checks, and outlier removal.

[0038] For any position in the structural anchor point ,from Select the one that is closest to the anchor point in the coordinate domain of the structure. Let the matching point be denoted as the nth matching point. The neighbors are The inverse squared distance kernel in this embodiment is... ;in, To prevent numerically stable terms with zero distance; the normalized weights are... .

[0039] The position is obtained by weighted interpolation of the displacement vectors of the nearest matching points. Corresponding position in semantically clear reference image ,Right now .

[0040] Under the assumption of smooth local deformation, the above process extends the discrete matching relationship into a continuous dense displacement field covering the coordinate domain of the structural anchor points. For regions with too few matching points or insufficient spatial coverage, the neighborhood range can be increased, and local affine fitting, radial basis functions, or thin plate splines can be used to supplement the data. If the estimation reliability is still lower than the preset threshold, the corresponding position is reverted to the original position in the structural anchor points to avoid unreliable long-distance extrapolation.

[0041] This embodiment uses the pre-defogging geometric anchor point. The upper part sets overlapping local windows according to the preset window size and step size. (Note: The first part is missing from the original text.) Each structure window is The corner points of the window are mapped to a semantically clear reference image using a dense displacement field. The corresponding region is determined by the mapped corner points. In each color channel Above, calculate the mean and standard deviation of the structure window and the reference region, and perform local statistical migration according to the following formula. ;in, Indicates the number after migration A window in the channel pixel values ​​on and These represent the channel mean and standard deviation, respectively. To prevent stable terms from having excessively small local standard deviations, This indicates element-wise multiplication.

[0042] The window step size is set to be smaller than the window size, allowing the same pixel to be covered by multiple windows. Let the covering pixel be... The set of window indices is The resulting aligned image for .

[0043] Overlapping averaging can reduce the block boundary discontinuities introduced by local statistical transformations. Besides mean-standard deviation transformation, appearance transfer can also employ histogram matching, local linear color transformation, whitening-coloring transformation, or learned local mapping, all of which preserve the geometric structure. Provided by, appearance statistics by The provided decoupling relationship.

[0044] Although overlapping local migration can significantly reduce block artifacts, matching errors, displacement interpolation errors, and local statistical differences can still cause problems. Color jumps, abnormal high-frequency details, or local structural distortions may occur. In this embodiment, these abnormalities are considered as mild degradation on a clear image and are repaired using a truncation-diffusion process.

[0045] Use an image encoder Will Mapped to the latent space, let the total diffusion time step be T, and the re-noise ratio be... The cutoff time step is Add noise to the latent variables up to the cutoff time step, i.e. ;in, The cumulative signal retention coefficient, Gaussian noise. Based on the original foggy image. As an image condition, from The reverse denoising process begins, and the decoder ultimately obtains the pseudo-clear label. .by As a conditional sustainable constraint on scene structure, truncating the sampling path avoids the complete back diffusion from causing a significant reconstruction of color and structure.

[0046] In this embodiment, , ,thus .when When set to 10, 20, and 30 respectively, the repair results are relative to The CIEDE2000 color difference can be 1.45, 2.56, and 4.87 respectively. The shorter truncation path can remove alignment artifacts while keeping color variations at a low level.

[0047] The Conditional Latent Diffusion Student Model Training Module compresses the processing power of the geometry-appearance decoupled teacher data construction pipeline into a single dehazing model. Training samples consist of original hazy images. With the corresponding pseudo-clear tag The teacher pipeline runs only during the offline data construction phase; student models can independently perform dehazing inference after training is complete.

[0048] In this embodiment, the student model uses a pre-trained latent diffusion model as the generation backbone, and sets conditional control branches corresponding to the backbone coding blocks and intermediate blocks. Original foggy image. Mapped to conditional latent variables by the image encoder and the noisy latent variables at the current diffusion time step. The conditional control branch is input after concatenation along the channel dimension. The outputs of each encoding stage and intermediate block of the conditional control branch are zero-initialized. Convolution generates multi-scale controlled features.

[0049] Let the bottleneck characteristics of the backbone network be... The control features generated by the intermediate block of the conditional control branch are Then the bottleneck feature is updated as follows In the backbone decoding stage, let the current decoding feature be... The skip connection feature corresponding to the encoding stage is Conditional control features are Then the decoded input is updated to Through the above process, low-level edges, spatial layout, and high-level semantics with fog input can continuously constrain clear image reconstruction at different scales. Zero-initialization convolution brings the control signal close to zero in the early stages of training, thereby avoiding disruption of the feature distribution of the pre-trained backbone.

[0050] This embodiment will use pseudo-clear labels. via encoder Mapping to clear latent variables For the diffusion time step of random sampling The forward noise addition process is The student noise prediction network is denoted as ,in, This represents the trainable parameters. The network uses... Time step and foggy image conditions Given the input, predict the added noise. The training loss is By minimizing the loss, the semantic reasoning and natural appearance capabilities provided by the image editing model, the geometric constraints provided by the structural anchors, and the natural image distribution constraints provided by the artifact restoration process are all incorporated into the student model parameters.

[0051] In this embodiment, the generated backbone uses pre-trained weights from Stable Diffusion v2.1 and keeps them frozen. The conditional control branches are initialized from the control branch weights of an existing diffusion dehazing model, and only the conditional control branches are updated during training. The training images are randomly cropped. Pixels, applying JPEG compression perturbation to foggy input with a certain probability; the optimizer uses AdamW, with a learning rate set to... The batch size is 32, the total training steps are 50,000, and the diffusion training time steps are set to 1,000. The above configuration is only an example; the generated backbone, trainable parameter range, optimizer, and training parameters can all be adjusted according to the deployment goals.

[0052] Physically perceptual spatial adaptive inference is executed on the trained student model without changing its training weights. Instead, during backdiffusion, it dynamically adjusts the generation guidance intensity at different locations based on the spatial distribution of transmittance in the input foggy image. Specifically: The foggy image to be processed Encoding as conditional representation From standard Gaussian latent variables Starting from the preset diffusion sampler, estimate step by step. During the sampling process, positive text conditions describing clear, natural scenes and negative text conditions describing degradation such as fog, blur, noise, and overexposure were used to obtain positive conditional noise predictions. and negative conditional noise prediction .

[0053] This embodiment uses a dark channel prior to estimate transmittance. For foggy images... Local dark channels are ;in, Represented in pixels A local window centered on the center. Indicates the color channel.

[0054] Atmospheric light can be estimated by selecting pixels with higher brightness from the high-value region of the dark channel in the original image. To avoid numerical anomalies caused by excessively small atmospheric light channel components, a positive lower bound is introduced. The transmittance is estimated to be ;in, To retain a small amount of haze and adjust the coefficient for estimated intensity, This is the lower bound of transmittance; the obtained Describe the degree to which the observed signal is preserved at different locations in the input image.

[0055] It should be noted that transmittance can also be obtained from color attenuation priors, depth estimation models, or learnable physical parameter networks.

[0056] In this embodiment, a spatially guided intensity map is constructed based on the estimated transmittance. ;in, This indicates the basic guiding strength in areas with high transmittance. This indicates the maximum additional guidance range.

[0057] when When it approaches 1, near The student model prioritizes preserving the already reliable structure and color in the input image; when When approaching the lower limit of transmittance, As the number of students increases, the student model makes more use of generative priors learned from pseudo-paired data to recover details obscured by fog.

[0058] It should be noted that the spatial guiding function is not limited to this linear form. In other embodiments, piecewise linear functions, power functions, exponential functions, or data-calibrated monotonic functions can also be used, as long as the relationship that the generating guiding intensity does not decrease when the transmittance decreases is satisfied.

[0059] The spatial guided intensity map S(x) is processed by the resampling operator. Adjust to latent variables Spatial resolution. At each back-diffusion time step, the modulated noise is predicted as... ;in, This represents the noisy prediction used to update the latent variables. This represents element-wise multiplication. The formula extends global scalar guidance to restore control over spatial variations, enabling the model to enhance generative completion only in regions where physically observed information is significantly diminished.

[0060] In one embodiment, , The diffusion sampling step count is 50. The spatial guidance map is calculated once before sampling begins and reused in all back-diffusion time steps; therefore, the additional calculations mainly include one transmittance estimation and one scale transformation.

[0061] In this embodiment, two real-world foggy image datasets, URHI and RTTS, are used. URHI includes various outdoor scenes such as streets, buildings, and natural landscapes, covering different fog concentrations and spatial distributions. It is mainly used for constructing pseudo-paired training data and training the student model. URHI initially contains approximately 4800 original foggy images. After removing images that overlap with RTTS and filtering out images with failed editing or insufficient clarity reference quality, 3417 images are retained. RTTS contains 4322 original foggy images, which are used as an independent test set and are not used for pseudo-paired data construction, student model training, or parameter selection to verify the generalization ability of this invention for unseen real-world fog distributions.

[0062] In the pseudo-pairing data construction stage, when constructing a dense displacement field from sparse matching points, the following selection is made: The nearest neighbor matching point; the re-noise ratio for truncation diffusion artifact repair is set to... This corresponds to the 10th time step out of 50 inference time steps. During training, the backbone is generated using a pre-trained latent diffusion model and kept frozen; only the conditional control branches are updated. The conditional control branches are initialized with the corresponding parameters of the pre-trained DiffDehaze model. The optimizer is AdamW with a learning rate of [missing value]. Batch size is 32, random cut size is The total training steps were 50,000, with a diffusion training time of 1,000 steps. JPEG compression perturbation was applied to foggy inputs with a probability of 0.5. Training was completed using eight NVIDIA A40 graphics processors.

[0063] The inference phase simultaneously inputs a foggy image condition, positive text conditions describing a clear and natural appearance, and negative text conditions describing fog, blur, noise, and compression artifacts. The backdiffusion sampling step count is set to 50, and the base strength of the physically-aware spatial adaptive guidance is [value missing]. The maximum additional strength is Therefore, the theoretical upper limit of the spatial guidance intensity is 4.0.

[0064] To verify the technical effectiveness of this embodiment, it was compared with representative codebook prior methods, domain adaptation methods, and diffusion dehazing methods on an independent real-world foggy image test set. Since real-world foggy images typically lack strictly paired ground truth values, referenceless image quality metrics such as MUSIQ, Q-Align, CLIPIQA+, MANIQA, QualiCLIP+, ARNIQA, and DeQA were used to evaluate the dehazing results in terms of sharpness, texture naturalness, color quality, semantic plausibility, and artifact suppression. Higher values ​​for each metric indicate better image quality.

[0065] The comparison methods cover different technical approaches: DCP and CAP are traditional methods based on artificial physical or statistical priors; MSBDN, DehazeFormer, and BioIR represent convolutional multi-scale restoration, Transformer dehazing, and biomimetic image restoration methods, respectively; RIDCP, IPC-Dehaze, CoA, and DiffDehaze represent codebook priors, iterative priors, domain adaptation, and diffusion generative methods in recent real-world dehazing, respectively. In addition to directly deployable dehazing methods, the experiments also showcase pseudo-clear labels generated by the offline teacher data construction pipeline of this invention. This is used to compare the degree to which student models distill and surpass teacher capabilities; the pseudo-clear label is for reference only and is not a baseline for independently deployable online dehazing.

[0066] RTTS is completely independent of data construction and model training; therefore, Table 1 reflects the generalization ability of this invention on unseen real fog distributions. This invention outperforms all deployable comparison methods across all seven metrics. Calculated using unrounded raw data, compared to the suboptimal baselines for each metric, MANIQA, CLIPIQA+, and QualiCLIP+ show improvements of approximately 19.8%, 15.2%, and 14.8%, respectively, with an average relative gain of approximately 13% across the seven metrics. Qualitative results indicate that traditional methods such as DCP and CAP are prone to fog residue and color anomalies in complex lighting or non-uniform dense fog areas; RIDCP and IPC-Dehaze can recover some local textures, but may produce color shifts or inconsistencies in detail when the real fog distribution deviates from its training domain; DiffDehaze has a more natural overall appearance, but may still exhibit fog residue or overly smoothed structures in extremely low visibility areas. This invention, through geometry-appearance decoupling supervision and spatial adaptive generation guidance, can enhance semantic details in dense fog areas while reducing unnecessary modifications to light fog and near-field structures.

[0067] Table 1. Comparison of no-reference image quality metrics for different methods on the RTTS independent test set.

[0068] On the RTTS test set not seen, the student model in this embodiment shows pseudo-clear labels relative to the teacher pipeline. All seven indicators improved, with MANIQA and Q-Align showing relative improvements of approximately 11.1% and 8.2% respectively. The average relative improvement across the seven indicators was approximately 6.8%.

[0069] The student model does not mechanically copy pseudo-labels sample by sample. The teacher pipeline consists of multiple interconnected steps, including image editing, cross-appearance matching, local transfer, and generative inpainting. Matching errors, local color anomalies, or editing artifacts in individual samples may accumulate progressively. The student model, on the other hand, learns a unified end-to-end restoration mapping from a large number of pseudo-paired samples, preserving the stable semantic and appearance patterns found in the teacher pipeline and mitigating occasional sample-level errors. Latent diffusion training further establishes clear image distribution constraints, while physical perception spatial adaptive guidance supplements details that the teacher pseudo-labels may not fully express in low-transmittance regions. Therefore, the distillation process achieves both teacher capability compression and distribution-level regularization and quality improvement of the teacher pseudo-labels.

[0070] Table 2. Comparison of pseudo-clear label quality under different data construction module settings.

[0071] After removing the pre-dehazing structural anchors, the contrast, texture sharpness, and local appearance differences between the original hazy image and the semantically clear reference increased, and the number of low-quality samples with fewer than 100 matching points increased to about 10 times that of the complete scheme; MUSIQ decreased from 65.61 to 61.45, and all other indicators also decreased, indicating that pre-dehazing not only improved the reliability of cross-appearance correspondence, but also reduced the pressure on subsequent artifact restoration to simultaneously bear dehazing and error correction.

[0072] When removing cross-appearance matching and dense displacement fields, and directly performing block statistical migration after scaling two images to the same size, Q-Align decreases from 3.79 to 3.49, and DeQA decreases from 3.54 to 3.29, indicating that explicit geometric correspondence is a necessary condition for ensuring spatial consistency of pseudo-labels. When removing truncated diffusion artifact repair and directly using local migration results as pseudo-labels, CLIPIQA+, QualiCLIP+, and DeQA decrease from 0.68, 0.63, and 3.54 to 0.62, 0.59, and 3.20, respectively, indicating that the truncated diffusion process can effectively suppress the block effect, high-frequency inconsistency, and local structural anomalies introduced by block migration. Table 2 demonstrates that pre-dehazing, explicit geometric matching, and generative artifact repair respectively undertake the functions of structural anchoring, correspondence correction, and natural image distribution projection, and the three together constitute a closed loop for generating high-quality pseudo-paired data.

[0073] As shown in Table 3, after removing the physically-aware spatial adaptive guidance, all six indicators decreased, with MUSIQ dropping from 67.61 to 65.95, demonstrating that allocating guidance intensity based on transmittance can effectively improve the dehazing quality of real-world scenes. Further, when the upper limit of the spatial guidance range, s=4.0, is used as a fixed scalar for the entire image, although the overall contrast and dehazing intensity in dense fog areas improve, over-enhancement, color changes, or structural modifications are prone to occur in light fog and foreground areas; fixing s=1.0 results in insufficient recovery of low-transmittance dense fog areas. This invention adopts... The spatial adaptive guidance enables low-transmittance regions to obtain stronger generative priors while high-transmittance regions maintain lower intervention, thereby forming a more stable balance between detail restoration and structural fidelity.

[0074] Table 3. Ablation results guided by physical perception space.

[0075] To verify whether improvements in quality metrics without a reference can be translated into human subjective preference, this embodiment organized 22 volunteers to conduct anonymous pairwise evaluations of RIDCP, IPC-Dehaze, CoA, DiffDehaze, and the present invention, resulting in 395 sets of comparison results. Participants rated the products on a scale of 1 to 5 across four dimensions: visibility, color and naturalness, detail retention, and visual comfort, and calculated their ELO rankings based on overall preference.

[0076] As shown in Table 4, the highest results were achieved across all four subjective dimensions and the ELO ranking. In pairwise comparisons with RIDCP, CoA, DiffDehaze, and IPC-Dehaze, the number of "win / match / lag" results for this embodiment were 30 / 4 / 7, 25 / 4 / 12, 24 / 6 / 10, and 19 / 7 / 16, respectively. The advantages in visibility and color naturalness are particularly evident, indicating that the data construction process can convey natural color and texture priors to the student model, while physical perception guidance can improve visibility in foggy areas and avoid uniform enhancement of reliable areas.

[0077] Table 4. Subjective Perception Scores and ELO Rankings Based on Different Methods

[0078] This embodiment utilizes original foggy images to construct pseudo-paired data with consistent structure and natural appearance; the student model effectively distills and outperforms the cascaded teacher pipeline on most metrics; explicit geometric correspondence and truncated diffusion inpainting can reduce spatial misalignment and local artifacts, respectively; transmittance-based spatial adaptive guidance enhances detail recovery in low-transmittance regions and preserves structure and color in high-transmittance regions. The above results validate the feasibility and synergistic effects of this embodiment's technical solution from multiple perspectives, including unseen scene generalization, training data quality, module independent contribution, and human subjective perception.

[0079] Example 2 Embodiment 2 of the present invention introduces an image dehazing system based on generative distillation and physical perception guidance.

[0080] like Figure 3 The image dehazing system shown includes: The acquisition module is configured to acquire the raw, foggy image. The generation module is configured to generate a semantically clear reference image and pre-dehazed geometric anchor points based on the acquired original hazy image; The extraction module is configured to extract cross-appearance matching points based on the generated semantically clear reference image and pre-dehazed geometric anchor points; The fusion module is configured to establish geometric correspondences between images based on the extracted cross-appearance matching points, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels. The adjustment module is configured to form pseudo-paired data based on the original foggy image and the image pseudo-clear labels, use the pseudo-paired data to train the conditional latent diffusion student model, perform physical perception based on the original foggy image, and construct a spatial guidance intensity map. The dehazing module is configured to spatially modulate the positive and negative conditional noise predictions based on the spatial guidance intensity map, obtain the dehazed image through backdiffusion sampling, and complete the image dehazing based on generative distillation and physical perception guidance.

[0081] The detailed steps are the same as those of the image dehazing method based on generative distillation and physical perception guidance provided in Example 1, and will not be repeated here.

[0082] Example 3 Embodiment 3 of the present invention provides a computer-readable storage medium.

[0083] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of an image dehazing method based on generative distillation and physical perception guidance as described in Embodiment 1 of the present invention.

[0084] The detailed steps are the same as those of the image dehazing method based on generative distillation and physical perception guidance provided in Example 1, and will not be repeated here.

[0085] Example 4 Embodiment 4 of the present invention provides an electronic device.

[0086] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in an image dehazing method based on generative distillation and physical perception guidance as described in Embodiment 1 of the present invention.

[0087] The detailed steps are the same as those of the image dehazing method based on generative distillation and physical perception guidance provided in Example 1, and will not be repeated here.

[0088] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0089] A computer program product includes software code, wherein the program in the software code performs the steps of an image dehazing method based on generative distillation and physical perception guidance as described in Embodiment 1 of the present invention.

[0090] The detailed steps are the same as those of the image dehazing method based on generative distillation and physical perception guidance provided in Example 1, and will not be repeated here.

[0091] 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 implemented 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. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0092] 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 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] 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.

[0094] 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.

[0095] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0096] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0097] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. An image dehazing method based on generative distillation and physical perception guidance, characterized in that, include: Obtain the original foggy image; Based on the acquired original foggy image, a semantically clear reference image and pre-dehazed geometric anchor points are generated. Based on the generated semantically clear reference image and pre-dehazed geometric anchor points, extract cross-appearance matching points; Based on the extracted cross-appearance matching points, establish the geometric correspondence between images, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels; Pseudo-paired data is formed based on the original foggy image and the pseudo-clear image labels. The conditional latent diffusion student model is trained using the pseudo-paired data. Physical perception is performed based on the original foggy image to construct a spatial guidance intensity map. Spatial modulation is performed on the positive and negative conditional noise predictions based on the spatial guidance intensity map, and the dehazed image is obtained by back diffusion sampling, thus completing the image dehazing based on generative distillation and physical perception guidance.

2. The image dehazing method based on generative distillation and physical perception guidance as described in claim 1, characterized in that, The original hazy image is processed using dehazing editing instructions to generate a semantically clear reference image. The dehazing editing instructions include at least removing the effects of fog and atmospheric scattering to enhance the texture and contrast of the visible area, maintaining the scene category, image aspect ratio, basic semantics of objects and original lighting context, and maintaining the original color context. The generated pre-dehazing geometric anchor points are consistent with the pixel coordinates, object boundaries, relative positions and scene layout of the original hazy image.

3. The image dehazing method based on generative distillation and physical perception guidance as described in claim 1, characterized in that, In the process of extracting cross-appearance matching points, semantically clear reference images and pre-dehazed geometric anchor points are both input into the cross-appearance image matching model to obtain cross-appearance matching points; the cross-appearance image matching model adopts the MatchAnything model.

4. The image dehazing method based on generative distillation and physical perception guidance as described in claim 1, characterized in that, In the process of obtaining the pseudo-clear label of the image, local windows that overlap each other are set on the pre-dehazed geometric anchor points according to the preset window size and step size. The corner points of the local windows are mapped onto the semantically clear reference image through the displacement field. The corresponding region of the semantically clear reference image is determined based on the mapped corner points. Local statistical migration across appearance matching points is performed in the determined region. The local statistical migration values ​​are overlapped and fused to obtain a fused aligned image; the truncated diffusion process is used to repair the obtained fused aligned image to obtain the image pseudo-clear label.

5. The image dehazing method based on generative distillation and physical perception guidance as described in claim 1, characterized in that, In the process of physical perception of the original foggy image, conditional backdiffusion sampling is performed on the original foggy image to obtain positive conditional noise prediction and negative conditional noise prediction of the image. Image transmittance is estimated based on the original foggy image to obtain a spatial guidance intensity map; spatial modulation is performed on the positive conditional noise prediction and the negative conditional noise prediction based on the spatial guidance intensity map to obtain a noise prediction used to update the latent variables.

6. The image dehazing method based on generative distillation and physical perception guidance as described in claim 1, characterized in that, Based on the original foggy image and the structure-preserving defogging model, pre-defogging geometric anchor points are generated; wherein the structure-preserving defogging model adopts a conditional diffusion-based defogging model or a geometry-based defogging model.

7. An image dehazing system based on generative distillation and physical perception guidance, characterized in that, include: The acquisition module is configured to acquire the raw, foggy image. The generation module is configured to generate a semantically clear reference image and pre-dehazed geometric anchor points based on the acquired original hazy image; The extraction module is configured to extract cross-appearance matching points based on the generated semantically clear reference image and pre-dehazed geometric anchor points; The fusion module is configured to establish geometric correspondences between images based on the extracted cross-appearance matching points, perform local appearance transfer and overlap fusion, and obtain image pseudo-clear labels. The adjustment module is configured to form pseudo-paired data based on the original foggy image and the image pseudo-clear labels, use the pseudo-paired data to train the conditional latent diffusion student model, perform physical perception based on the original foggy image, and construct a spatial guidance intensity map. The dehazing module is configured to spatially modulate the positive and negative conditional noise predictions based on the spatial guidance intensity map, obtain the dehazed image through backdiffusion sampling, and complete the image dehazing based on generative distillation and physical perception guidance.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of an image dehazing method based on generative distillation and physical perception guidance as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of an image dehazing method based on generative distillation and physical perception guidance as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of an image dehazing method based on generative distillation and physical perception guidance as described in any one of claims 1-6.

Images