Fog removal method and device for remote sensing image, equipment, storage medium and program product

Abstract

The application relates to the technical field of image processing, and discloses a method for removing fog from a remote sensing image, which comprises the following steps: based on an input remote sensing image, an implicit atmospheric condition field is constructed, and amplitude information is recovered and phase information is refined based on the remote sensing image according to the implicit atmospheric condition field, so that a target remote sensing image is obtained. The target remote sensing image is subjected to fog removal, so that a remote sensing image after fog removal is obtained. According to the implicit atmospheric condition field, a remote sensing image with fog is generated based on the remote sensing image after fog removal. A de-fogging model is trained according to the loss between the input remote sensing image and the remote sensing image with fog, and the de-fogging model is used to remove fog from the remote sensing image. The method can train the de-fogging model to learn an implicit frequency field without relying on atmospheric prior data and training data, so that the de-fogging model can effectively remove fog from the remote sensing image, and the de-fogging effect of the remote sensing image is improved. The application also discloses a device and equipment for removing fog from a remote sensing image, a storage medium and a program product.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus, device, storage medium, and program product for defogging remote sensing images. Background Technology

[0002] Remote sensing imagery plays a crucial role in fields such as environmental monitoring, ground observation, and intelligent aviation navigation. To eliminate atmospheric interference, restore the true information in remote sensing images, and ensure image usability and data reliability, defogging is necessary.

[0003] In related technologies, the following two methods are mainly used to remove fog from remote sensing images: one is to build a fog removal model based on prior atmospheric data, and the other is to use a data-driven deep learning method to train the fog removal model using synthesized paired data as training data. The fog removal model obtained by either method can remove fog from remote sensing images.

[0004] However, under complex or non-uniform haze conditions, the defogging models obtained from related technologies are difficult to effectively defog remote sensing images, thus reducing the defogging effect of remote sensing images. Summary of the Invention

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0006] This application provides a method, apparatus, device, storage medium, and program product for defogging remote sensing images. It can train a defogging model to learn an implicit frequency field without relying on prior atmospheric data and training data, enabling the defogging model to effectively defog remote sensing images, thereby improving the defogging effect of remote sensing images.

[0007] In a first aspect, embodiments of this application provide a method for dehazing remote sensing images, comprising: Based on the input remote sensing images, an implicit atmospheric condition field is constructed; Based on the implicit atmospheric condition field, the amplitude information and phase information are recovered and refined from the input remote sensing image to obtain the target remote sensing image; Dehazing is performed on the target remote sensing image to obtain a dehazed remote sensing image; Based on the implicit atmospheric condition field, a foggy remote sensing image is generated on the basis of the defogging remote sensing image. A defogging model is trained based on the loss between the input remote sensing image and the foggy remote sensing image. The defogging model is used to defog remote sensing images that need to be defoged.

[0008] Optionally, based on the input remote sensing image, an implicit atmospheric condition field is constructed, including: jointly modeling haze density, illumination imbalance and scattering direction through pseudo-exposure generation and phase flow estimation.

[0009] Optionally, haze density, illumination imbalance, and scattering direction are jointly modeled by pseudo-exposure generation and phase flow estimation, including: performing multi-level gamma correction on the input haze image to generate a pseudo-exposure image sequence; calculating the brightness difference superposition based on the pseudo-exposure image sequence and combining it with phase flow estimation of scattering direction information; and fusing brightness difference, phase flow, and position coding through a lightweight coding network to output haze density, attenuation channel direction, and scattering direction.

[0010] Optionally, based on the implicit atmospheric condition field, amplitude information and phase information are recovered from the input remote sensing image to obtain the target remote sensing image. This includes: converting the input remote sensing image to the frequency domain to extract the amplitude spectrum and phase spectrum; enhancing the amplitude spectrum through a frequency diffusion branch and applying physical consistency constraints to the phase in the phase spectrum to obtain an enhanced remote sensing image; dehazing the input remote sensing image through a physical dehazing branch to obtain a pre-dehazed remote sensing image; and fusing the enhanced remote sensing image and the pre-dehazed remote sensing image through a gated fusion mechanism to obtain the target remote sensing image.

[0011] Optionally, the target remote sensing image is dehazed to obtain a dehazed remote sensing image, including: performing multi-scale residual extraction on the target remote sensing image to obtain a dehazed remote sensing image.

[0012] Optionally, based on the implicit atmospheric condition field, a foggy remote sensing image is generated from the defogging remote sensing image, including: controlling the refogging of the defogging image based on the parameters in the implicit atmospheric condition field to generate a foggy remote sensing image.

[0013] Secondly, embodiments of this application provide a defogging device for remote sensing images, comprising: The implicit atmospheric condition field construction module is configured to construct an implicit atmospheric condition field based on the input remote sensing image. The frequency domain optimization network module is configured to recover amplitude information and refine phase information based on the input remote sensing image according to the implicit atmospheric condition field, so as to obtain the target remote sensing image; The dynamic residual aggregation module is configured to defog the target remote sensing image to obtain a defog-free remote sensing image, and generate a foggy remote sensing image based on the defog-free remote sensing image according to the implicit atmospheric condition field. The model training module is configured to train a dehazing model based on the loss between the input remote sensing image and the foggy remote sensing image. The image dehazing module is configured to use a dehazing model to dehaze the remote sensing image to be dehazed.

[0014] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing program instructions, wherein the processor is configured to execute the defogging method for remote sensing images in the first aspect when running the program instructions.

[0015] Fourthly, embodiments of this application provide a storage medium storing program instructions, wherein the program instructions, when executed, perform the defogging method for remote sensing images as described in the first aspect.

[0016] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the defogging method for remote sensing images in the first aspect.

[0017] The defogging method, apparatus, device, storage medium, and program product for remote sensing images provided in this application can achieve the following technical effects: This application provides a novel frequency domain perspective for implicit haze modeling, addressing the problem of spatially uneven distribution of haze in remote sensing images. By employing an implicit atmospheric condition field and a dual-branch frequency optimization network, spatially adaptive and physically consistent image restoration is achieved. Furthermore, controllable fuzziness and cyclic consistency constraints eliminate dependence on paired training data. Thus, the defogging model can be trained to learn the implicit frequency field without relying on prior atmospheric data or training data. This enables the defogging model to effectively defog remote sensing images under complex or non-uniform haze conditions, thereby improving the defogging effect of remote sensing images.

[0018] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are considered similar elements. The drawings do not constitute a limitation of scale, and wherein: Figure 1 This is a diagram illustrating the effect of exchanging amplitude and phase according to an embodiment of this application; Figure 2 This is a network architecture diagram of PIF-Net provided in an embodiment of this application; Figure 3 This is a flowchart of a method for defogging remote sensing images provided in an embodiment of this application; Figure 4 This is a visualization of the UAV dataset from the experiment provided in this application embodiment; Figure 5This is a visualization of the EuroSAT dataset used in the experiment provided in this application embodiment; Figure 6 This is a visualization of the SIHI-WHU dataset from the experiments provided in this application embodiment; Figure 7 These are graphs showing the quantitative analysis results of different datasets in the experiments provided in this application embodiment; Figure 8 This is a visualization of the RICE dataset from the experiments provided in this application embodiment; Figure 9 This is a result diagram of the ablation experiment provided in the embodiments of this application; Figure 10 This is a visualization of the ablation experiment provided in the embodiments of this application; Figure 11 This is a schematic diagram of a defogging device for remote sensing images provided in an embodiment of this application; Figure 12 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] Remote sensing images play a vital role in environmental monitoring, ground observation, and intelligent aviation navigation. However, in real-world scenarios, atmospheric degradation often introduces haze effects, obscuring key information in remote sensing images. Unlike the relatively uniform distribution of haze in natural scenes, haze in remote sensing images exhibits spatial non-uniformity and temporal coupling. Its distribution is influenced by topographic elevation, surface reflectivity, and changes in illumination, resulting in highly complex and diverse degradation patterns.

[0021] In related technologies, remote sensing image dehazing methods are mainly divided into two categories: one is model-based methods that rely on atmospheric scattering priors, and the other is data-driven deep learning methods that learn representations from large-scale datasets.

[0022] While model-based methods exhibit a degree of physical interpretability, their defogging performance often deteriorates under non-uniform haze conditions due to overly idealized assumptions in atmospheric models. Data-driven deep learning methods, although achieving significant improvements in visual quality, have limited generalization to real-world scenes because the training data is primarily synthesized rather than captured under natural haze conditions. Therefore, under complex or non-uniform haze conditions, the defogging models obtained from the two aforementioned methods struggle to effectively defog remote sensing images, thus reducing the overall defogging effect.

[0023] It is worth mentioning that among the related technologies, there is a diffusion-based defogging method in the data-driven deep learning approach. Although this method improves the perceived realism, it is still difficult to effectively handle the degradation of haze caused by inherent spatial variations in complex remote sensing images, that is, it cannot guarantee the defogging effect of remote sensing images.

[0024] Specifically, regarding model-based methods, early research on remote sensing image dehazing primarily relied on atmospheric scattering to construct dehazing models, which attempted to explicitly estimate transmittance and global atmospheric light. Classical priors such as dark channel priors and color attenuation priors provided physically interpretable formulas and improved performance in uniform fog. Later research further introduced improved priors, including boundary constraint and context regularization models, nonlocal dehazing, and guided image filtering models, which improved edge preservation and reduced halos. Recently, some studies have re-examined model-based methods in the context of remote sensing and UAV imagery, combining adaptive atmospheric priors and illumination compensation. For example, by addressing the non-uniform scattering problem, dehazing and underwater image restoration have been improved. By combining physical models with image translation networks, a balance between realism and interpretability has been struck in dehazing tasks. Furthermore, by introducing a saliency-guided parallel learning mechanism, the degradation of haze in images can be effectively mitigated. However, these methods still rely on simplified physical assumptions, limiting their scalability under dynamic haze conditions.

[0025] For data-driven deep learning methods, with the development of deep learning, data-driven approaches have become the mainstream for dehazing remote sensing images. These methods directly learn haze-related representations from the data without requiring explicit physical modeling. Based on different network structures and learning paradigms, they can be divided into diffusion-based dehazing methods and other data-driven deep learning methods.

[0026] For diffusion-based dehazing methods, diffusion models have garnered widespread attention in various imaging tasks due to their powerful generative priors and excellent restoration capabilities. In the field of remote sensing image dehazing, by introducing DTPM (Dynamic Texture Perception Model), explicit texture priors are injected into the diffusion process, reducing randomness and improving perception quality. Furthermore, a Fourier prior-guided diffusion framework has been proposed, which utilizes amplitude and phase cues to allow null-radius joint low-light enhancement and deblurring, demonstrating good transfer to fog-related degradation. Additionally, RSHazeDiff, a unified Fourier perception diffusion model for remote sensing dehazing, has been developed, which can better preserve color and structure under non-uniform haze. Although diffusion-based methods achieve perceptual realism and high-quality restoration, they typically incur high computational costs and require careful noise scheduling, limiting their applicability in high-resolution remote sensing scenarios.

[0027] For other data-driven deep learning methods, beyond diffusion, dehazing research in a broader data-driven environment involves architectural design, spectral geometric cues, and supervision. For example, DEA-Net is proposed, which uses detail-enhanced convolutions and content-guided attention to preserve texture and edges. MB-TaylorFormer is also introduced, which achieves efficient high-resolution restoration through Taylor unfolding approximate self-attention. MITNet is proposed, which decouples the amplitude and structure-refining phase of haze attenuation in the frequency domain and further enhances this through deep-assisted collaboration. UCL-Dehaze, an unpaired contrastive adversarial pipeline, is also developed to align sharp and blurred regions. It is evident that most data-driven methods still face challenges when applied to non-uniform haze in remote sensing images, failing to adapt to variations in haze intensity and scattering direction in complex terrain. Furthermore, purely data-driven frameworks often lack explicit interpretability and physical awareness, necessitating novel designs to connect frequency domain modeling with physically inspired representation learning.

[0028] Therefore, embodiments of this application provide a method, apparatus, device, storage medium, and program product for defogging remote sensing images. Combined with... Figure 1 As shown, inspired by Fourier difference, the amplitude and phase components exhibit different physical response modes during haze degradation. Specifically, the amplitude mainly reflects the radiative attenuation caused by haze density, while the phase determines the structure and geometric features of the image. When the amplitude and phase are exchanged between blurred and sharp images, images A and C correspond to haze intensity, while images B and D maintain structural integrity. This reveals two fundamental dimensions of haze degradation: luminosity attenuation and structural distortion, which can be effectively decoupled and modeled separately in the frequency domain.

[0029] Based on this, this application proposes a unified physical-implied frequency field network (PIF-Net) dehazing framework, which can achieve physical consistency and spatially adaptive dehazing of remote sensing images. The framework first constructs an implicit atmospheric condition field, simulating haze density, illumination imbalance, and scattering direction through pseudo-exposure generation and phase flow estimation. Then, a frequency domain optimization network is designed to collaboratively recover amplitude and phase information through two complementary branches (a frequency diffusion branch and a physics-based dehazing branch), and maintains physical consistency through gated fusion. Finally, a dynamic residual aggregation module is introduced to achieve self-supervised learning under non-pair conditions, enabling the dehazing model to effectively generalize without ground truth annotations. That is, without relying on prior atmospheric data and training data, the dehazing model is trained to learn the implicit frequency field, enabling it to effectively dehaze remote sensing images, thereby improving the dehazing effect.

[0030] To provide a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this application. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0031] The terms "first," "second," etc., used in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. Unless otherwise stated, the term "a plurality of" means two or more.

[0032] In this embodiment, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B. The term "and / or" describes an association relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or, A and B. The term "correspondence" can refer to an association or binding relationship; A corresponding to B means that there is an association or binding relationship between A and B.

[0033] Combination Figure 2As shown in the embodiments of this application, a network architecture applicable to a remote sensing image dehazing method is provided. This network architecture can be the PIF-Net dehazing framework. The goal of PIF-Net is to recover clear remote sensing images from hazy observations affected by spatially inhomogeneous haze conditions. To this end, PIF-Net learns to construct physically implicit frequency fields that encode radiative attenuation and structural distortions caused by haze. Specifically, the input is defined as a single blurred image, from which the network estimates the implicit atmospheric representation and recovers a physically consistent haze-free image. PIF-Net comprises a three-level architecture: (a) an Implicit Atmospheric Condition Field (IACF) that simulates haze density, illumination imbalance, and scattering direction through pseudo-exposure generation and phase flow estimation; (b) a Frequency-domain Optimization Network (FON) that refines amplitude and phase through frequency diffusion and physically guided dehazing branches; and (c) a Dynamic Residual Aggregation with Controllable Re-hazing (DRA) module for self-supervised training of the dehazing model. Among these, for example... Figure 2 As shown, Conv×1 represents 1×1 convolution, PFT represents phase function transform, Conv×3×3 represents 3×3 convolution, HTF represents harmonic transfer function, Subpix represents subpixel (convolution), FFT (fast Fourier transformation) represents fast Fourier transform, IFFT (inverse fast Fourier transform) represents inverse fast Fourier transform, MLP (multilayer perceptron) represents multilayer perceptron, Sigmoid represents Sigmoid activation function, ReLU represents ReLU activation function, Conv×3×3 represents 3×3 convolution, L cyc This represents the loss of cycle consistency.

[0034] Combination Figure 3 As shown in the figure, this application provides a method for dehazing remote sensing images. This method can be applied to the PIF-Net framework of the above embodiments. The method includes the following steps: S31, based on the input remote sensing image, construct the implicit atmospheric condition field.

[0035] S32, based on the implicit atmospheric condition field, recovers the amplitude information and refines the phase information on the input remote sensing image to obtain the target remote sensing image.

[0036] S33, perform dehazing on the target remote sensing image to obtain a dehazed remote sensing image.

[0037] S34, Based on the implicit atmospheric condition field, generate a foggy remote sensing image on the basis of the defogging remote sensing image.

[0038] S35. Based on the loss between the input remote sensing image and the foggy remote sensing image, a defogging model is trained.

[0039] S36. Using a defogging model, defogging is performed on the remote sensing image to be defogging.

[0040] The defogging method for remote sensing images provided in this application offers a novel frequency domain perspective for implicit haze modeling, addressing the problem of spatially uneven distribution of haze in remote sensing images. Through an implicit atmospheric condition field and a dual-branch frequency optimization network, spatially adaptive and physically consistent image restoration is achieved, and controllable fuzziness and cyclic consistency constraints eliminate dependence on paired training data. Thus, the defogging model can be trained to learn the implicit frequency field without relying on prior atmospheric data and training data, enabling the defogging model to effectively defog remote sensing images under complex or non-uniform haze conditions, thereby improving the defogging effect of remote sensing images.

[0041] The following is a detailed explanation of steps S31 to S36.

[0042] Optionally, in step S31 above, an implicit atmospheric condition field is constructed based on the input remote sensing image, including: jointly modeling haze density, illumination imbalance, and scattering direction through pseudo-exposure generation and phase flow estimation. Specifically, multi-level gamma correction is performed on the input haze image to generate a pseudo-exposure image sequence. The brightness difference superposition is calculated based on the pseudo-exposure image sequence, and the scattering direction information is combined with phase flow estimation. A lightweight coding network is used to fuse the brightness difference, phase flow, and position coding to output the haze density, attenuation channel direction, and scattering direction.

[0043] In this implementation, haze exhibits dynamic and spatially varying characteristics in non-uniform remote sensing images. To characterize the spatially non-uniform haze intensity without a time series, multiple gamma values ​​are introduced to generate pseudo-exposure versions of the haze input image. By observing the differences between the pseudo-exposures, the dehazing model can perceive changes in haze density and attenuation level. The formula for the synthesized pseudo-multiradiative sequence is as follows:

[0044] In the above formula, express False exposure image, This indicates the corresponding exposure gain. Indicates the gamma correction factor. Indicates pixel position, This represents the color channel index. Based on this, the following brightness difference superposition is constructed, with the specific formula as follows:

[0045] In the above formula, This represents the corresponding tensor. This represents the brightness difference between the (k+1)th frame and the kth frame (i.e., the brightness change between two adjacent frames); k=1 and K−1 represent the frame index range, from the 1st frame to the K−1th frame, corresponding to the brightness difference of K−1 adjacent frames (a total of K frames). express The dimensions / data types, where, Representing a real number type, H is the height of the image (number of pixels), W is the width of the image (number of pixels), and (K−1) corresponds to the number of channels for the brightness difference (i.e., containing brightness difference information of K−1 adjacent frames).

[0046] Next, a time-independent phase flow is designed to represent the directional scattering behavior of fog particles. The specific formulas for the pixel-domain triples of amplitude and direction are as follows:

[0047] In the above formula, Represents the triplet features of the pixel domain. Indicates pixel position, and Let represent the direction confidence and the principal orientation angle, respectively. Based on this, a physically interpretable scattering direction prior can be derived from a single pseudo-exposure frame. Then, the obtained intensity variation and direction information are unified into a lightweight network to simulate the implicit atmospheric field. The specific formula is as follows:

[0048] In the above formula, , , These represent the density, attenuation channel direction, and scattering direction of the haze, respectively, with I representing the original image. Represents the set of brightness differences, and FFF represents the pixel-domain triplet feature. Indicates a hybrid encoder. This indicates the corresponding positional encoding. This implementation is achieved through channel-level connectivity, thereby enabling effective functional integration across multiple modes.

[0049] Optionally, in step S32 above, based on the implicit atmospheric condition field, amplitude information is recovered and phase information is refined from the input remote sensing image to obtain the target remote sensing image. This includes: converting the input remote sensing image to the frequency domain to extract the amplitude spectrum and phase spectrum; enhancing the amplitude spectrum through a frequency diffusion branch and applying physical consistency constraints to the phase in the phase spectrum to obtain an enhanced remote sensing image; dehazing the input remote sensing image through a physical dehazing branch to obtain a preliminarily dehazed remote sensing image; and fusing the enhanced remote sensing image and the preliminarily dehazed remote sensing image through a gated fusion mechanism to obtain the target remote sensing image.

[0050] In this implementation, under the implicit atmospheric field prior constraints provided by IACF, the amplitude is explicitly recovered and the phase is slightly refined in the frequency domain, while a physically consistent gated fusion spatial branch is constructed. This enhances image contrast and suppresses color distortion. Specifically, IACF consists of two branches: a frequency domain diffusion branch and a physically-based dehazing branch. The triple components from IACF are used to assign pixel weights to each pseudo-exposure image as the joint input of the two branches, as shown in the following formula:

[0051] In the above formula, This represents the fused remote sensing image of the target. K represents the pixel position, and K represents the total number of pseudo-exposure images (i.e., the number of images participating in the fusion). This represents the pixel value at pixel x in the k-th pseudo-exposure image. This represents the weight score of the k-th pseudo-exposure image at pixel x. This represents the weight score of the j-th pseudo-exposure image at pixel position x.

[0052] For the frequency-domain diffusion branch, the input image is first transformed into the frequency domain. The amplitude is then amplified within a specific direction and frequency band using the IACF prior, while allowing only minimal phase perturbations to maintain structural consistency. Specifically, the amplitude and phase components are decomposed using Fourier transform, as shown in the following formula:

[0053]

[0054] In the above formula, Indicates the amplitude component, Represents the phase component, For frequency domain coordinates, For channel index, For amplitude spectrum, For phase spectrum, This represents a two-dimensional Fast Fourier Transform.

[0055] Then, amplitude diffusion is performed in the logarithmic domain, specifically to force physical consistency of the phase as follows:

[0056] In the above formula, This represents the corrected (physically consistent) frequency domain phase, where K represents the wavenumber vector in the frequency domain. This represents the original frequency domain phase (distinguished by channel c). This represents the truncation function. This represents the upper bound of the permissible phase deviation. The function representing the calculation of phase correction amount, Local weight score (or local feature).

[0057] Finally, the output of the frequency domain branch is obtained through inverse Fourier transform, as shown in the following formula:

[0058] In the above formula, This indicates the output of the frequency domain branch. This represents the two-dimensional inverse fast Fourier transform. This represents the corrected frequency domain amplitude. This represents the element-wise multiplication operator. This represents the complex exponential term corresponding to the corrected frequency domain phase.

[0059] For the physics-based dehazing branch, a lightweight physics-based module is used to approximate the inverse process of the Atmospheric Scattering Model (ASM), achieving global brightness correction and haze removal. The global atmospheric light estimation formula is as follows:

[0060] In the above formula, This indicates a global average pooling operation. This represents the estimated global brightness for each channel. The physically-based dehazing operation formula is as follows:

[0061] In the above formula, This represents the pixel value at pixel x and channel c in the dehazed image. This represents the pixel value at pixel x and channel c of the input remote sensing image. and These represent the transmission diagram and atmospheric light, respectively. This represents the lower bound of transmission. Finally, an enhanced image (i.e., a remote sensing image of the target) is generated through an adaptive decision-making mechanism.

[0062] Optionally, in step S33 above, dehazing the target remote sensing image to obtain a dehazed remote sensing image includes: performing multi-scale residual extraction on the target remote sensing image to obtain a dehazed remote sensing image.

[0063] In this implementation, since it is difficult to obtain real-world remote sensing images with corresponding ground truth (GT), a GT-free supervised training strategy is provided. Stable and clear images (i.e., dehazed remote sensing images) are obtained through multi-scale residual aggregation. This process creates a self-supervised cyclic constraint between the original and synthesized images. The formula for the spatial fusion process is as follows:

[0064] In the above formula, This represents a clear image after multi-scale residual aggregation processing. Represents a remote sensing image of a target. This represents the lightweight residual component. This indicates a gating mapping.

[0065] Optionally, in step S34 above, generating a foggy remote sensing image based on the defogging remote sensing image according to the implicit atmospheric condition field includes: performing controllable refogging on the defogging image based on the parameters in the implicit atmospheric condition field to generate a foggy remote sensing image.

[0066] In this implementation, haze is controllably synthesized on a clear image using the field prior provided by IAF, thereby generating a guiding haze counterpart. The aim of this implementation is to preserve the restoration in dense haze areas while enhancing structural details in texture-rich areas. To achieve controllable re-hazing, the clear image is reintroduced into the haze under the guidance of IACF, ensuring that its frequency domain characteristics remain close to those of the real haze image. The specific formula is as follows:

[0067] In the above formula, This represents the synthesized haze image output after controlled re-fogging processing. Atmospheric Light, estimated by IACF, represents the global or local ambient light intensity at location (x, c) in the image and is a key parameter for haze formation. The input image is a clear image, which is the haze-free image obtained after dehazing and is also the basis for rehashing. This represents the transmittance obtained by IAF guidance.

[0068] To further illustrate the beneficial effects of the defogging method for remote sensing images provided in the embodiments of this application, an experimental scheme is provided to verify the defogging method for remote sensing images provided in the embodiments of this application, as follows: Experimental Datasets: The UAV, EuroSAT, SRI-WHU, and RICE datasets were used to evaluate the defogging performance of the methods. The UAV dataset is a partially publicly available, fuzzy remote sensing dataset used to evaluate the defogging performance of various methods. The EuroSAT dataset, derived from Sentinel-2 satellite imagery, covers 13 spectral bands, includes subsets of ten different scenes, a total of 27,000 annotated and georeferenced images, and 2,000–3,000 RS images for each scene category, each 64×64 pixels in size. This dataset is helpful for testing the robustness of various haze removal methods in different scene environments. The RICE dataset has two subsets: RICE1 and RICE2, both of which are real-world remote sensing datasets. The RICE1 dataset contains 500 pairs of 512×512 resolution images collected from Google Earth.

[0069] Baseline method of the experiment: The PIF-Net method of the present application embodiment is compared with image dehazing methods such as CGID, DNMGDT, PMT, RIDCP, SSID, BCDP, GIFM, and DEA-Ne.

[0070] Experimental Details: To ensure fairness in comparing the defogging method provided in this application with other defogging methods, the results reported by PIF-Net and all comparison methods were performed on the same device and under the same configuration settings. For baseline methods, their source code was used and the reproduced performance was reported. For deep learning-based methods, the ADAM optimizer with default parameters and an initial learning rate of 0.0001 was used. This experiment was conducted in a PyCharm 2022 environment using Python 3.7 and CUDA 11.6. For traditional methods, this experiment was implemented using MATLAB R2020a.

[0071] Experimental metrics: Four commonly used image quality assessment metrics were employed: average gradient (AG), edge intensity (EI), patch-based contrast quality index (PCQI), and information entropy (IE) to quantitatively evaluate the dehazing effects of various methods. Average gradient (AG), often referred to as sharpness, describes the sharpness of an image and indicates the contrast level of image details and texture variations. A higher average gradient value indicates a sharper image. A higher edge intensity score indicates clearer texture recovery. PCQI is a method for evaluating contrast variations, primarily focusing on local contrast quality maps. Its significant advantage lies in its applicability to local image patches, creating quality maps that capture spatial variations and providing valuable information about local quality changes. For the IE metric, a higher score indicates richer information contained in the image.

[0072] Qualitative comparison of experiments: The PIF-Net provided in this application consistently produces visually superior dehazing results on the UAV, EuroSAT, and SIHI datasets. Combined with... Figures 4 to 6 As shown, the experiments provide visualizations on the UAV dataset, the EuroSAT dataset, and the SIHI-WHU dataset, respectively. Figures 4 to 6 In the table, a, b, c, d, e, f, g, h, i, and j correspond to dehazing methods such as CGID, DNMGDT, PMT, RIDCP, SSID, BCDP, GIFM, and DEA-Ne, respectively. For example, Figure 4 As shown in the visualization on the UAV dataset, traditional prior-based methods such as CGID and DNMGDT cannot handle dense and spatially varied haze, often resulting in residual fog and overly smooth surfaces. Figure 5 As shown, learning-based methods, such as PMT and GIFM, are prone to color differences and radiation imbalances due to the complex spectral distribution of vegetation and urban textures. Figure 6 As shown, large-scale haze caused by topographical variations leads to models such as SSID and DEA-Net amplifying high-frequency noise or blurring structural edges. In contrast, PIF-Net effectively combines frequency domain amplitude-phase interactions with physically guided priors to achieve adaptive recovery, preserving fine details and maintaining global color consistency. These results highlight the robustness and interpretability of PIF-Net under various haze distributions and imaging conditions.

[0073] Quantitative comparison of experiments: combined Figure 7As shown, the experiments provide quantitative analysis tables on three datasets. PIF-Net, provided in this embodiment, achieves state-of-the-art quantitative performance across all datasets, significantly outperforming prior- and data-driven baselines. On the UAV dataset, traditional methods (such as CGID and DNMGDT) fail under dense, non-uniform haze due to their reliance on simplified atmospheric priors, resulting in weak contrast and low edge strength. On the EuroSAT dataset, learning-based methods such as PMT and SSID exhibit color bias and structure loss due to overfitting to synthetic data, failing to generalize to complex multispectral terrain. On the SIRI-WHU dataset, models such as RIDCP and DEA-Net tend to over-enhance brightness while suppressing texture, leading to unnatural visual artifacts. In contrast, PIF-Net effectively combines physically implicit frequency fields with amplitude and phase recovery, achieving a balanced enhancement of global illumination and fine structure. This dehazing model achieves sharpness, contrast, and radiometric consistency, demonstrating strong robustness and physical interpretability across various remote sensing domains.

[0074] Differential detection in experiments: In order to better evaluate the performance of the defogging method provided in the embodiments of this application, combined with Figure 8 As shown, the experiment presents the dehazing results and corresponding error maps on the RICE dataset. The error maps reveal that cooler-toned areas indicate lower reconstruction errors, demonstrating that the dehazing method provided in this embodiment maintains good consistency and edge sharpness when removing spatially uneven haze. In contrast, other methods are prone to local over-enhancement or texture suppression, while the dehazing model in this embodiment effectively achieves a balance between global contrast restoration and local detail preservation. Furthermore, as... Figure 8 The quantitative results of PSNR / SSIM shown verify the advantages of the defogging method in this embodiment under complex and non-uniform haze conditions in remote sensing images.

[0075] Ablation Experiments: To verify the effectiveness of each component in the PIF-Net provided in this application, a comprehensive ablation study was conducted on three remote sensing datasets. Specifically, the IACF, FON, and DRA modules were successively removed to evaluate their contributions. Figure 9 As shown, the experiment provides an ablation experiment analysis table. Figure 9 The results show that removing IAF leads to a significant degradation in all metrics, especially EI and AG, indicating that modeling spatially varying haze density and illumination imbalance is crucial for accurate reconstruction. Removing FON significantly decreases PCQI and IE values, demonstrating that joint amplitude and phase refinement in the frequency domain effectively enhances structural and radiometric consistency. Excluding DRA results in a decrease in EI and PCQI, indicating that a controllable heavy fog mechanism improves the stability of self-supervised learning and texture preservation. Figure 10 As shown, the experiment provides a visualization of the ablation experiment, such as... Figure 10 The visual comparisons shown further support these findings: without IACF, the images remain generally dull; without FON, local contrast and spectral balance are compromised; without DRA, fine textures become overly smooth. The full PIF-Net recovers clearer structure and natural color distribution across all datasets, exhibiting the highest quantitative and perceptual performance.

[0076] Experiments show that the defogging model provided in this application outperforms other defogging methods on multiple datasets, providing a new paradigm for physically guided frequency domain recovery of remote sensing images.

[0077] The following describes the product embodiments provided in this application.

[0078] Combination Figure 11 As shown, this application provides a defogging device 110 for remote sensing images. The defogging device 110 includes: an implicit atmospheric condition field construction module 111, a frequency domain optimization network module 112, a dynamic residual aggregation module 113, a model training module 114, and an image defogging module 115.

[0079] The implicit atmospheric condition field construction module 111 is configured to construct an implicit atmospheric condition field based on the input remote sensing image. The frequency domain optimization network module 112 is configured to recover amplitude information and refine phase information based on the input remote sensing image according to the implicit atmospheric condition field, so as to obtain the target remote sensing image; The dynamic residual aggregation module 113 is configured to defog the target remote sensing image to obtain a defog-free remote sensing image, and generate a foggy remote sensing image based on the defog-free remote sensing image according to the implicit atmospheric condition field. The model training module 114 is configured to train a dehazing model based on the loss between the input remote sensing image and the foggy remote sensing image. Image defogging module 115 is configured to use a defogging model to defog the remote sensing image to be defogged.

[0080] Optionally, the implicit atmospheric condition field construction module 111, when constructing the implicit atmospheric condition field based on the input remote sensing image, is specifically used to: jointly model haze density, illumination imbalance and scattering direction through pseudo-exposure generation and phase flow estimation.

[0081] Optionally, the implicit atmospheric condition field construction module 111, when jointly modeling haze density, illumination imbalance, and scattering direction through pseudo-exposure generation and phase flow estimation, is specifically used for: performing multi-level gamma correction on the input haze image to generate a pseudo-exposure image sequence; calculating the brightness difference superposition based on the pseudo-exposure image sequence and combining it with phase flow estimation of scattering direction information; and fusing brightness difference, phase flow, and position encoding through a lightweight coding network to output haze density, attenuation channel direction, and scattering direction.

[0082] Optionally, the frequency domain optimization network module 112, in the process of recovering amplitude information and refining phase information based on the input remote sensing image according to the implicit atmospheric condition field to obtain the target remote sensing image, specifically performs the following: converting the input remote sensing image to the frequency domain to extract the amplitude spectrum and phase spectrum; enhancing the amplitude spectrum through a frequency diffusion branch and applying physical consistency constraints to the phase in the phase spectrum to obtain an enhanced remote sensing image; dehazing the input remote sensing image through a physical dehazing branch to obtain a preliminarily dehazed remote sensing image; and fusing the enhanced remote sensing image and the preliminarily dehazed remote sensing image through a gated fusion mechanism to obtain the target remote sensing image.

[0083] Optionally, the dynamic residual aggregation module 113 is specifically used in the process of dehazing the target remote sensing image to obtain a dehazed remote sensing image to: extract multi-scale residuals from the target remote sensing image to obtain a dehazed remote sensing image.

[0084] Optionally, in the process of generating a foggy remote sensing image based on a defogging remote sensing image according to the implicit atmospheric condition field, the dynamic residual aggregation module 113 is specifically used to: perform controllable refogging on the defogging image based on the parameters in the implicit atmospheric condition field to generate a foggy remote sensing image.

[0085] The defogging device for remote sensing images provided in this application embodiment is used to implement the defogging method for remote sensing images in the foregoing embodiments. The specific implementation method and beneficial effects can be referred to the foregoing embodiments, and will not be repeated here.

[0086] Combination Figure 12 As shown, this application embodiment provides an electronic device 120, including a processor 121 and a memory 122. Optionally, the electronic device 120 may further include a communication interface 123 and a bus 124. The processor 121, memory 122, and communication interface 123 can communicate with each other via the bus 124. The communication interface 123 can be used for information transmission. The processor 121 can call logical instructions in the memory 122 to execute the defogging method for remote sensing images described in the above embodiment.

[0087] Furthermore, the logic instructions in the aforementioned memory 122 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0088] The memory 122, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this application. The processor 121 executes functional applications and data processing by running the program instructions / modules stored in the memory 122, thereby implementing the defogging method for remote sensing images in the above embodiments.

[0089] The memory 122 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 122 may include high-speed random access memory and may also include non-volatile memory.

[0090] This application provides a storage medium storing computer-executable instructions, which are configured to execute the defogging method for remote sensing images described in the above embodiments.

[0091] The aforementioned storage medium can be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0092] The technical solutions of this application embodiment can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this application embodiment. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0093] This application provides a computer program product, including a computer program that, when executed by a processor, implements the defogging method for remote sensing images described in the above embodiments.

[0094] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0096] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions, and operations of the system, method, and computer program product according to the embodiments of this application. In this respect, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the figures. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the figures, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for dehazing remote sensing images, characterized in that, include: Based on the input remote sensing images, an implicit atmospheric condition field is constructed; Based on the implicit atmospheric condition field, the amplitude information and phase information are recovered and refined on the input remote sensing image to obtain the target remote sensing image; The target remote sensing image is dehazed to obtain a dehazed remote sensing image; Based on the implicit atmospheric condition field, a foggy remote sensing image is generated on the basis of the defogging remote sensing image. A defogging model is trained based on the loss between the input remote sensing image and the foggy remote sensing image. The aforementioned defogging model is used to defog the remote sensing image to be defogged.

2. The method according to claim 1, characterized in that, The implicit atmospheric condition field is constructed based on the input remote sensing image, including: Haze density, illumination imbalance, and scattering direction are jointly modeled by pseudo-exposure generation and phase flow estimation.

3. The method according to claim 2, characterized in that, The method of jointly modeling haze density, illumination imbalance, and scattering direction through pseudo-exposure generation and phase flow estimation includes: Multi-level gamma correction is applied to the input haze image to generate a pseudo-exposure image sequence; The brightness difference superposition is calculated based on the pseudo-exposure image sequence, and the scattering direction information is estimated by combining the phase flow. By fusing brightness difference, phase flow, and position coding through a lightweight coding network, the output of haze density, attenuation channel direction, and scattering direction is obtained.

4. The method according to claim 1, characterized in that, Based on the implicit atmospheric condition field, amplitude information and phase information are recovered from the input remote sensing image to obtain the target remote sensing image, including: The input remote sensing image is converted to the frequency domain to extract the amplitude spectrum and phase spectrum; The amplitude spectrum is enhanced by frequency diffusion branching, and the phase in the phase spectrum is physically constrained to obtain an enhanced remote sensing image. The input remote sensing image is dehazed through a physical dehazing branch to obtain a preliminarily dehazed remote sensing image; The enhanced remote sensing image and the preliminarily dehazed remote sensing image are fused using a gating fusion mechanism to obtain the target remote sensing image.

5. The method according to claim 1, characterized in that, The process of dehazing the target remote sensing image to obtain a dehazed remote sensing image includes: Multi-scale residual extraction is performed on the target remote sensing image to obtain the dehazed remote sensing image.

6. The method according to claim 1, characterized in that, Based on the implicit atmospheric condition field, a foggy remote sensing image is generated on the basis of the defogging remote sensing image, including: Based on the parameters in the implicit atmospheric condition field, the defogging image is subjected to controllable refogging to generate a foggy remote sensing image.

7. A defogging device for remote sensing images, characterized in that, include: The implicit atmospheric condition field construction module is configured to construct an implicit atmospheric condition field based on the input remote sensing image. The frequency domain optimization network module is configured to recover amplitude information and refine phase information based on the input remote sensing image according to the implicit atmospheric condition field, so as to obtain the target remote sensing image; The dynamic residual aggregation module is configured to defog the target remote sensing image to obtain a defog-free remote sensing image, and generate a foggy remote sensing image based on the defog-free remote sensing image according to the implicit atmospheric condition field. The model training module is configured to train a defogging model based on the loss between the input remote sensing image and the foggy remote sensing image. The image defogging module is configured to use the defogging model to defog the remote sensing image to be defoged.

8. An electronic device comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the defogging method for remote sensing images as described in any one of claims 1 to 6 when running the program instructions.

9. A storage medium storing program instructions, characterized in that, When the program instructions are executed, they perform the defogging method for remote sensing images as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The system includes a computer program that, when executed by a processor, implements the defogging method for remote sensing images as described in any one of claims 1 to 6.

Images