Remote sensing image defogging method, device, equipment, storage medium and program product

Abstract

The application discloses a remote sensing image defogging method and device, equipment, storage medium and program product, and relates to the technical field of image processing. The method comprises the following steps: acquiring a remote sensing image in fog, and performing preprocessing on the remote sensing image in fog based on an atmospheric scattering model to generate a preprocessed image; performing gamma correction on the preprocessed image to generate a plurality of exposure adjustment images, and generating a global enhancement image based on the plurality of exposure adjustment images; performing enhancement processing on a local image block in the preprocessed image based on an enhancement control factor to generate a local enhancement image; and fusing the global enhancement image and the local enhancement image to obtain a defogging remote sensing image. Through the above method, high-frequency edge details and low-frequency degradation areas are processed respectively, so that more effective defogging is realized.

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 dehazing remote sensing images. Background Technology

[0002] Remote sensing (RS) images can objectively and in real-time depict the Earth's surface landscape, serving as a crucial means of perceiving and understanding geographic information. However, due to the strong scattering and absorption of light by particulate matter in the atmosphere over long distances, remote sensing images are easily affected by haze during acquisition, frequently suffering from image blurring and loss of detail. True geographic details are thus obscured, leading to a significant deterioration in image quality. Therefore, achieving image dehazing is of great importance.

[0003] Traditional methods for image dehazing typically employ image restoration, image enhancement, or deep learning approaches. However, these methods require substantial computational resources. Furthermore, haze in remote sensing images often exhibits significant spatial inhomogeneity, accompanied by a complex coexistence of high- and low-frequency features. Traditional dehazing methods primarily focus on mitigating the haze effect from a single or single-scale perspective, often resulting in the loss of high-frequency details in distant areas or heavy reliance on high-quality training data. In practical applications, these methods suffer from both detail loss and high computational complexity. Summary of the Invention

[0004] 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.

[0005] This application provides a method, apparatus, device, storage medium, and program product for dehazing remote sensing images, which respectively process high-frequency edge details and low-frequency degradation areas, thereby achieving more effective dehazing.

[0006] In a first aspect, embodiments of this application provide a method for dehazing remote sensing images, comprising: acquiring a hazy remote sensing image, and preprocessing the hazy remote sensing image based on an atmospheric scattering model to generate a preprocessed image; performing gamma correction on the preprocessed image to generate multiple exposure adjustment images, and generating a global enhanced image based on the multiple exposure adjustment images; enhancing local image patches in the preprocessed image based on an enhancement control factor to generate a local enhanced image; and fusing the global enhanced image and the local enhanced image to obtain a dehazed remote sensing image.

[0007] Optionally, performing gamma correction on the preprocessed image to generate multiple exposure adjustment images, and generating a global enhanced image based on the multiple exposure adjustment images, includes: performing gamma correction on the preprocessed image based on multiple gamma values ​​to generate multiple exposure adjustment images; extracting the global and local components of each exposure adjustment image; and fusing the global and local components of the multiple exposure adjustment images to generate a global enhanced image.

[0008] Optionally, fusing the global and local components of multiple exposure-adjusted images includes: constructing a global weight map and a local weight map; and fusing the global and local components of multiple exposure-adjusted images based on the global weight map and the local weight map.

[0009] Optionally, the step of enhancing local image patches in the preprocessed image based on an enhancement control factor to generate a locally enhanced image includes: converting the preprocessed image from the RGB color space to the CIELAB color space to calculate the mean and variance of local image patches in the luminance channel in the CIELAB color space; determining the enhancement control factor based on the ratio of the global variance to the variance of the local image patches, and enhancing the contrast of the local image patches based on the enhancement control factor to obtain an enhanced image; using a guided filter to suppress noise in the enhanced image to obtain a denoised image; and converting the denoised image from the CIELAB color space to the RGB color space to obtain the locally enhanced image.

[0010] Optionally, the step of preprocessing the foggy remote sensing image based on the atmospheric scattering model to generate a preprocessed image includes: obtaining atmospheric light parameters from the foggy remote sensing image; and preprocessing the foggy remote sensing image using the atmospheric scattering model based on the atmospheric light parameters to generate the preprocessed image.

[0011] Optionally, the step of extracting the global and local components of each exposure-adjusted image includes: filtering the luminance component in each exposure-adjusted image using a guided filter to obtain the global component.

[0012] Secondly, embodiments of this application provide a remote sensing image dehazing device, comprising: a preprocessing module for acquiring a foggy remote sensing image and preprocessing the foggy remote sensing image based on an atmospheric scattering model to generate a preprocessed image; a global enhancement module for performing gamma correction on the preprocessed image to generate multiple exposure adjustment images, and generating a global enhanced image based on the multiple exposure adjustment images; a local enhancement module for enhancing local image patches in the preprocessed image based on an enhancement control factor to generate a local enhanced image; and a fusion module for fusing the global enhanced image and the local enhanced image to obtain a dehazed remote sensing image.

[0013] 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 method described in the first aspect or any alternative manner when running the program instructions.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the method described in the first aspect or any of the alternative methods.

[0015] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect or any of the optional methods.

[0016] The remote sensing image dehazing method provided in this application preprocesses a foggy remote sensing image to generate a preprocessed image. Then, gamma correction is applied to the preprocessed image to generate multiple exposure-adjusted images, and a global enhanced image is generated based on these. Simultaneously, local image patches in the preprocessed image are enhanced based on enhancement control factors to generate local enhanced images. Finally, the global and local enhanced images are fused to obtain a dehazed remote sensing image. This approach enhances global high-frequency information and local low-frequency details separately, resulting in both a global and a local enhanced image. Finally, the enhanced global and local enhanced images are aggregated, fusing complementary high and low-frequency information to obtain a clear and structurally coherent high-quality dehazed remote sensing image.

[0017] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0018] 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 an overall architecture diagram of AFAM provided in the embodiments of this application; Figure 2 A flowchart illustrating a remote sensing image dehazing method provided in this application embodiment; Figure 3 A schematic diagram of the exposure result after gamma correction provided in an embodiment of this application; Figure 4 Visual illustrations of different control factor selections provided in the embodiments of this application; Figure 5 A schematic diagram illustrating the fusion processing results provided in an embodiment of this application; Figure 6 A schematic diagram of a remote sensing image dehazing device provided in an embodiment of this application; Figure 7 A visual comparison diagram of the UAV dataset provided in the embodiments of this application; Figure 8 Quantitative analysis tables on three datasets provided in this application embodiment; Figure 9 Provided for the embodiments of this application Schematic diagram; Figure 10 A visual comparison diagram of the SIRI-WHU dataset provided in the embodiments of this application; Figure 11 Schematic diagrams showing the enhanced details provided for embodiments of this application; Figure 12 This is a schematic diagram of ablation experiment results provided in an embodiment of this application; Figure 13 Quantitative analysis table of ablation experiments provided for embodiments of this application; Figure 14 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0021] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0023] Remote sensing (RS) images can objectively and in real-time depict the Earth's surface landscape, serving as a crucial means of perceiving and understanding geographic information. However, remote sensing images are easily affected by haze during acquisition, obscuring true geographic details and significantly degrading image quality. Consequently, image dehazing has received increasing attention.

[0024] The degradation process of the haze effect can be described by the Atmospheric Scattering Model (ASM), as follows: I(x) = J(x)t(x) + A(1−t(x)) Where I(x) represents a foggy image, J(x) is an ideal fog-free image, t(x) represents the transmittance of the medium, and A represents the atmospheric light component.

[0025] Image restoration-based dehazing methods typically restore hazy images by estimating t(x), but their performance is often limited by inaccurate parameter estimation, resulting in incomplete dehazing. Image enhancement-based dehazing methods improve image visibility by adjusting contrast and applying exposure fusion, but they often struggle to preserve high-frequency details and accurate color representation.

[0026] In recent years, deep learning-based methods have introduced bidirectional transformations of fog-clear-fog and clear-fog-clear to ensure content consistency. However, these methods require significant computational resources and perform poorly in preserving low-frequency texture details. Their main limitation is their inability to simultaneously capture global high-frequency features and local low-frequency details, and their difficulty in establishing complementary relationships between the two frequency domains.

[0027] Specifically, traditional image restoration methods utilize prior knowledge to estimate transmittance and reverse the degradation process, thereby obtaining a clear, haze-free image. For example, Dark Channel Prior (DCP) refines a coarse transmittance map by estimating scene depth information and combining it with fast guided filtering, resulting in a relatively clear dehazing effect. However, blurring often remains in the final restored image. Color Attenuation Prior (CAP) estimates scene sharpness by constructing a linear model, but due to the discontinuity of scene structure in remote sensing images, it is prone to introducing artifacts and halo effects.

[0028] Traditional image enhancement methods improve image visibility through color correction and contrast enhancement. For example, RMSP introduces a coefficient of variation to characterize the discreteness of input samples and implements adaptive gradient image enhancement based on a traditional retinal model. WTSR proposed a remote sensing image enhancement method that integrates wavelet transform and sparse representation, which can simultaneously enhance global spatial details and local structure to obtain high-resolution images, but may also lead to image overexposure. In summary, although image enhancement methods can improve image quality by suppressing or enhancing specific feature information, they are insufficient to fully address the problems of color inaccuracy and overexposure in remote sensing images because they rely solely on the input image itself.

[0029] Furthermore, deep learning methods generate sharp images by learning the mapping between foggy and fog-free images. PAB-Net proposes an end-to-end dehazing model that combines parallel spatial and channel attention modules, thus addressing the shortcomings of CNNs in capturing spatial and channel features and achieving more accurate results. However, traditional remote sensing image dehazing methods are still prone to over-enhancement, color distortion, and artifacts.

[0030] In general, existing dehazing methods mainly focus on reducing the haze effect from a single or single-scale perspective. However, haze in remote sensing images often exhibits significant spatial inhomogeneity, accompanied by the complex coexistence of high and low frequency features, which places higher demands on detail restoration and multi-scale consistency.

[0031] Based on this, embodiments of this application provide a method, apparatus, device, storage medium, and program product for dehazing remote sensing images, which respectively process high-frequency edge details and low-frequency degradation areas, thereby achieving more effective dehazing. The executing entity of this application is an electronic device, such as a server, laptop, desktop computer, mobile phone, tablet computer, etc., and embodiments of this application are not limited thereto.

[0032] Figure 1 The overall architecture diagram of AFAM provided in the embodiments of this application, combined with Figure 1As shown, this embodiment proposes a framework for the Advantage Feature Aggregation Method (AFAM), which includes two main stages: an enhancement stage, which enhances global high-frequency information and local low-frequency details to obtain a globally enhanced image and a locally enhanced image; and an aggregation stage, which performs advantage feature aggregation on the enhanced representation, fuses high- and low-frequency complementary information, and reconstructs a clear and structurally coherent dehazed remote sensing image.

[0033] Specifically, this framework enhances structure and texture perception by fully utilizing multi-frequency information. The preprocessed image is input into a dual-branch architecture, modeling high-frequency and low-frequency dominant features in parallel. In one branch, a multi-frequency exposure enhancement mechanism is employed, using an adaptive weight map to guide the selective fusion of multiple exposure levels, thereby preserving high-frequency edge details and improving global contrast. In the other branch, an adaptive frequency compensation mapping strategy is used to focus on restoring brightness and details in low-frequency degraded areas. Unlike sequential processing, these two enhancement branches are executed in parallel. Finally, a dual-domain frequency fusion module integrates the complementary features of the two branches to reconstruct a clear and structurally consistent dehazed remote sensing image.

[0034] The following are definitions of terms: Input image (I): The input image (I), i.e., the fogged remote sensing image; Output image (F): The output image (F), i.e., the defoggy remote sensing image; AFA: Advantage Feature Aggregation; MFEE: Multi-Frequency Exposure Enhancement; DCP: Dark Channel Prior; AFCM: Adaptive Frequency Compensation Mapping; Global Weight: Global Weight Map; Detail Weight: Local Weight Map; Exposure image: The exposed image; Mean Map: The mean map; Local P: Local Patch; Enhancement: Enhancement; Filter: Filtering; High-frequency feature pyramid: High-frequency feature pyramid; Low-frequency feature pyramid: Low-frequency feature pyramid.

[0035] Figure 2 A flowchart illustrating a remote sensing image dehazing method provided in an embodiment of this application. (Combined with...) Figure 2As shown in the figure, this application provides a method for dehazing remote sensing images, which includes the following steps: S101: Acquire foggy remote sensing images and preprocess them based on an atmospheric scattering model to generate preprocessed images; S102: Perform gamma correction on the preprocessed image to generate multiple exposure-adjusted images, and generate a global enhancement image based on the multiple exposure-adjusted images; S103: Enhance local image patches in the preprocessed image based on enhancement control factors to generate locally enhanced images; S104: Fuse the global enhanced image and the local enhanced image to obtain the dehazed remote sensing image.

[0036] The remote sensing image dehazing method provided in this application preprocesses a foggy remote sensing image to generate a preprocessed image. Then, gamma correction is applied to the preprocessed image to generate multiple exposure-adjusted images, and a global enhanced image is generated based on these. Simultaneously, local image patches in the preprocessed image are enhanced based on enhancement control factors to generate local enhanced images. Finally, the global and local enhanced images are fused to obtain a dehazed remote sensing image. This approach enhances global high-frequency information and local low-frequency details separately, resulting in both a global and a local enhanced image. Finally, the enhanced global and local enhanced images are aggregated, fusing complementary high and low-frequency information to obtain a clear and structurally coherent high-quality dehazed remote sensing image.

[0037] Optionally, in step S101 above, preprocessing the foggy remote sensing image based on the atmospheric scattering model to generate a preprocessed image includes: obtaining atmospheric light parameters from the foggy remote sensing image, and then preprocessing the foggy remote sensing image based on the atmospheric light parameters using the atmospheric scattering model to generate a preprocessed image.

[0038] In this embodiment, inspired by the dark channel prior (DCP), key parameters are first extracted from the original image (i.e., the foggy remote sensing image) to estimate atmospheric light A, and a preprocessed image is generated based on the atmospheric scattering model, as shown in the following expression:

[0039] in, This represents the preprocessed image. Represents the original image. Indicates atmospheric light value, Indicates the transmittance of the medium. This represents the lower limit threshold of transmittance.

[0040] Optionally, in step S102 above, gamma correction is performed on the preprocessed image to generate multiple exposure adjustment images, and a global enhancement image is generated based on the multiple exposure adjustment images. This includes: performing gamma correction on the preprocessed image based on multiple gamma values ​​to generate multiple exposure adjustment images; then, extracting the global and local components of each exposure adjustment image; and then fusing the global and local components of the multiple exposure adjustment images to generate a global enhancement image.

[0041] Optionally, in the above embodiments, extracting the global and local components of each exposure adjustment image includes: filtering the luminance component in each exposure adjustment image using a guided filter to obtain the global component.

[0042] It should be noted that even after haze removal preprocessing, some haze remains in the remote sensing image, along with issues of low saturation and low brightness. Therefore, it is necessary to enhance the contrast and detail of the preprocessed image. This embodiment achieves global haze removal through a series of gamma corrections. Gamma correction can enhance the contrast between darker and brighter areas, but may lead to potential loss of detail in underexposed or overexposed areas. Therefore, inspired by MEF (Mean Exponentially Filtered Image), this embodiment uses a series of gamma value adjustments to change the image exposure, avoiding the loss of regional details and combining the local advantages of multiple exposures. First, the image intensity is globally modified through a power function transformation, as shown in the following formula:

[0043] in, Indicates the preprocessed image. Indicates the gamma value. This represents the scaling factor.

[0044] like Figure 3 As shown, Figure 3 This is a schematic diagram of the exposure result after gamma correction processing provided in an embodiment of this application; where (a) is the original image, and (b), (c), (d), and (e) are the exposure-adjusted images after processing with γ values ​​of 1.2, 2.0, 4.0, and 8.0, respectively. Figure 3 As shown in (b)-(e), power curve stretching can enhance the contrast in underexposed images and preserve more details. By selecting a series of gamma values, a set of images with adjusted exposure can be obtained. The global contrast enhancement strategy used in this embodiment relies on the Laplacian pyramid decomposition of overexposed images, sequentially extracting local structural details from artificially generated multi-exposure image sequences and merging dominant regions. By applying mean filtering to the luminance component, the global component can be directly obtained. Furthermore, to eliminate halos and artifacts caused by edge blurring, a guided filter can be introduced to obtain smooth global components, the expression of which is:

[0045] in, Represents the global component of the k-th image. Indicates a pilot filter. This represents the luminance component of the k-th image.

[0046] After multiple experiments, this embodiment fuses four manually adjusted exposure images (exposure adjusted images with γ values ​​of 1.2, 2, 4, and 8). After filtering the image brightness, local components can be directly estimated, as shown in the following expression:

[0047] in, Represents the local components of the k-th image. This represents the original brightness value of the k-th image. This represents the global component of the k-th image.

[0048] Optionally, in the above embodiments, fusing the global and local components of multiple exposure adjustment images includes: constructing a global weight map and a local weight map; and fusing the global and local components of multiple exposure adjustment images based on the global weight map and the local weight map.

[0049] In this embodiment, after calculating the global and local components of each gamma-corrected image (i.e., the exposure-adjusted image), it is further necessary to construct a corresponding weight map for the multi-exposure image fusion process. A general weight function based on exposure features is then used. Defined as: ; in, Indicates the weight value. Indicates the exposure characteristics of the image. Indicates the expected value. It represents the standard deviation.

[0050] According to this general weight function, when the exposure feature close to its expected value At this point, the calculated weights reach their maximum value. If some pixel values ​​are too far from the optimal value, the weights approach zero. For each location in the image, the global weight map... and local weight map Calculated using the following formula: ; ; in, Indicates the position of the k-th image. global weights This represents the global component at that location. This represents the average brightness value at that location. Indicates the adjustment parameter. The standard deviation of the global weights; Let represent the local weight of the k-th image at position (x, y). Characteristics representing local components, The standard deviation of the local weights.

[0051] In this embodiment, the optimal exposure constant can be set to 0.5, and the average brightness value is obtained by convolving the brightness component with a 7×7 average filter. After estimating and normalizing the weight mapping of all input images, the global and local components of multiple gamma-corrected images are fused to obtain the final fused image, expressed by the formula: ; in, This represents the merged image. This represents the global weight map of the k-th image. Represents the global component of the k-th image. Indicates the enhancement coefficient. This represents the local weight map of the k-th image. This represents the local component of the k-th image.

[0052] Gamma correction can reduce the brightness of blurred areas without losing content detail. However, gamma correction is a global contrast enhancement operation, often neglecting local detail. To compensate for this, a locally contrast-enhanced version of the initial dehazed image can be calculated by appropriately enhancing the high-frequency components of the luminance channel.

[0053] Optionally, in step S103 above, enhancing local image patches in the preprocessed image based on the enhancement control factor to generate a locally enhanced image includes: converting the preprocessed image from the RGB color space to the CIELAB color space to calculate the mean and variance of the local image patches in the luminance channel in the CIELAB color space; determining the enhancement control factor based on the ratio of the global variance to the variance of the local image patches, and enhancing the contrast of the local image patches based on the enhancement control factor to obtain the enhanced image; using a guided filter to suppress noise in the enhanced image to obtain a denoised image; and converting the denoised image from the CIELAB color space to the RGB color space to obtain the locally enhanced image.

[0054] In this embodiment, local image patch operations are introduced to enhance local contrast, preserving more fine texture components while achieving faster processing speed. First, the preprocessed image is converted from the RGB color space to the CIELAB color space to facilitate the extraction of target contours and recovery of fine textures from the luminance channel L. Second, assuming that the local image patch is a local integral image of the luminance channel L, the mean and variance of each local image patch are calculated using integral and square integral mappings, as shown in the following formulas: ; ; in, Let P represent the mean of the image patch, x represent the x-coordinate of the image patch, y represent the y-coordinate of the image patch, w represent the width of the image patch, h represent the height of the image patch, and N represent the total number of pixels in the image patch. This represents the brightness value of the pixel at coordinates (ij). This represents the variance of image patch P.

[0055] The above calculation strategy utilizes integral images to calculate pixel values ​​and the sum of squares of local image patches, reducing computational complexity while maintaining enhancement performance. Since the mean can be considered as a low-frequency component of a local image patch, the difference between the input image brightness value and the mean is used as a high-frequency component to enhance the image patch, as shown below:

[0056] in, This indicates that within image block P, it is located at... The brightness value after pixel enhancement. To enhance control factors, This represents the mean of image patch P.

[0057] Figure 4 Visual effect diagrams of different control factor selections provided in embodiments of this application, wherein (a) is a preprocessed image, and (b) is... The effect diagram with a value of 0.1 is shown in (c). The effect diagram with a value of 0.4 is shown in (d). The result of the value being 0.9, (e) is... =2.0 effect diagram, (f) is =4.5, (g) is the effect diagram. The effect diagram is 7.0, and (h) is the filtered noise diagram, as shown below. Figure 4 As shown, this embodiment presents visualization results under different selections of enhancement control factors. From Figure 4 In (b)-(g), it can be observed that, as As the pixel density increases, fine textures in the image become more apparent, but pixel overflow and oversaturation occur. Conversely, as the pixel density increases... The reduction in size makes the enhanced image smoother, which often leads to more residual haze and a loss of finer details. Therefore, choosing a suitable... Obtaining the optimal local contrast enhancement image is crucial. To achieve adaptive control of the enhancement level, this embodiment redefines the enhancement control factor as the variance of the entire image (i.e., the global variance). ) and local image patch variance (i.e. The ratio of ) and the formula for enhancement processing are rewritten as:

[0058] in, This indicates that within image block P, it is located at... The brightness value after pixel enhancement. To enhance control factors, This represents the variance of the entire image. Represents the variance of a local image patch. This represents the brightness value of the pixel at coordinates (ij). This represents the mean of image patch P. To adjust the threshold.

[0059] In this formula, This is used to keep the enhancement control factor within the normal enhancement range and prevent over-enhancement of the image. Furthermore, since noise may be introduced into the enhanced local image patches, this embodiment introduces a guided filter into the enhanced local image patches to suppress noise during the enhancement process, which can be expressed as: ; in, This represents the image after guided filtering. This represents the normalized guidance image. and This represents the linear parameters of the guided filter. Indicated by A local image patch centered on the image.

[0060] In this embodiment, a guidance image with less noise can be obtained by normalization processing (i.e., dividing the difference between the local enhanced image and its maximum value by the difference between the maximum and minimum values ​​of the local image patch). The optimal parameters can be obtained by minimizing the difference between the improved local image patch and the enhanced local image patch using a loss function. and The corresponding cost function Defined as: ; in, Represents a local neighborhood. and This represents the linear parameters of the guided filter. This represents the normalized guidance image. This represents the image after guided filtering.

[0061] The final denoised version of the enhanced image is as follows Figure 4 As shown in (h). Finally, the filtered image is converted from the CIELAB color space to the RGB color space to obtain a locally enhanced image, which helps to reduce the loss of local details caused by scattering.

[0062] Optionally, in the above embodiments, fusing the global enhanced image and the local enhanced image to obtain the dehazed remote sensing image includes: performing Gaussian pyramid decomposition on the global enhanced image and the local enhanced image respectively to obtain high-frequency dominant features and low-frequency dominant features; fusing the high-frequency dominant features and the low-frequency dominant features to obtain the dehazed remote sensing image.

[0063] Unlike traditional fusion methods, the fusion strategy proposed in this embodiment can maximize the fusion of high-frequency and low-frequency dominant features between globally and locally contrast-enhanced images, thereby producing clear remote sensing (RS) images. First, a discrete Gaussian filter is used to convolve the input image in both the horizontal and vertical directions, achieving two-dimensional smoothing of the image and removing noise while preserving key structural information. Mathematically, and Let $\mathbf$ represent the global contrast-enhanced image and the local contrast-enhanced image to be fused, respectively. A pyramid representation is constructed by low-pass filtering and downsampling these images. The approximate low-frequency component of the pyramid plane is obtained by the following equation:

[0064] in, This represents the approximate low-frequency component after merging the t-th layer of the pyramid. This represents the low-pass filtering and downsampling operation at layer t. This represents the global contrast-enhanced image of layer t. This represents the local contrast-enhanced image of layer t. express The approximate low-frequency component at layer t express Approximate low-frequency components in layer t.

[0065] In practice, the number of pyramid layers t depends on the size of the image and directly affects the visual quality of the blended image. and These represent images sampled at the pyramid level for the global enhancement image and the local enhancement image, respectively. and This represents the low-frequency components of the global and local enhanced images at layer t. Figure 5 (c) and (g) show the approximate low-frequency components of the globally contrast-enhanced image and the locally contrast-enhanced image, respectively. Figure 5 This is a schematic diagram of the fusion processing result provided in the embodiments of this application. The diagram shows the result of each core step in the framework from left to right. Among them, (a) is the original image, (b) is the global contrast-enhanced image, (c) is the approximate low-frequency component of the global contrast-enhanced image, (d) is the approximate high-frequency component of the global contrast-enhanced image, (e) is the initial dehazed image, (f) is the local contrast-enhanced image, (g) is the approximate low-frequency component of the local contrast-enhanced image, (h) is the approximate high-frequency component of the local contrast-enhanced image, and (i) is the final result image that fuses the advantages of high and low frequency features.

[0066] Based on the low-frequency components in the above formula, texture details can be enhanced by calculating high-frequency coefficients. Specifically, the high-frequency coefficient at each location is defined as the larger of the high-frequency components calculated from the two input images, expressed as:

[0067] in, This represents the high-frequency coefficients of the t-th level of the pyramid. This represents the global contrast-enhanced image of layer t. This represents the local contrast-enhanced image of layer t. and This represents the low-frequency components of the global and local enhanced images at layer t. This represents the minimum value.

[0068] Selecting and storing high-frequency coefficients according to the above formula can preserve the high-frequency and low-frequency dominant features of both the globally and locally enhanced images. Extensive experiments show that selecting a pyramid level of 3 achieves a better balance between fusion quality and processing speed. In summary, the enhanced dehazed image is obtained by the following formula:

[0069] in, This indicates the position of the final output dehazed image. Pixel value at that location, This represents the low-frequency components after fusion at layer t. This represents the high-frequency coefficients of the t-th layer of the pyramid.

[0070] This embodiment proposes an AFAM method for dehazing remote sensing images. By aggregating complementary advantages of high- and low-frequency features, it achieves efficient and robust dehazing. The multi-frequency exposure enhancement mechanism and adaptive frequency compensation mapping strategy respectively address high-frequency edge details and low-frequency degradation regions, thus achieving more effective dehazing. The advantageous feature aggregation strategy is used to fuse complementary features from high-frequency and low-frequency enhanced images, ultimately achieving structurally clear and visually coherent dehazed image reconstruction.

[0071] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0072] Figure 6 This is a schematic diagram of a remote sensing image dehazing device provided in an embodiment of this application. The device 600 includes: a preprocessing module 601, a global enhancement module 602, a local enhancement module 603, and a fusion module 604.

[0073] The preprocessing module 601 is used to acquire a foggy remote sensing image and preprocess the foggy remote sensing image based on an atmospheric scattering model to generate a preprocessed image. The global enhancement module 602 is used to perform gamma correction on the preprocessed image and generate multiple exposure adjustment images to generate a global enhancement image based on the multiple exposure adjustment images; The local enhancement module 603 is used to enhance local image patches in the preprocessed image based on enhancement control factors to generate a locally enhanced image; The fusion module 604 is used to fuse the global enhanced image and the local enhanced image to obtain the dehazed remote sensing image.

[0074] In one feasible implementation, the global enhancement module 602 performs gamma correction on the preprocessed image to generate multiple exposure adjustment images. When generating a global enhanced image based on the multiple exposure adjustment images, it performs gamma correction on the preprocessed image based on multiple gamma values ​​to generate multiple exposure adjustment images; extracts the global component and local component of each exposure adjustment image; and fuses the global component and local component of the multiple exposure adjustment images to generate a global enhanced image.

[0075] In one feasible implementation, the global enhancement module 602 is used to construct a global weight map and a local weight map when fusing the global and local components of multiple exposure-adjusted images; based on the global weight map and the local weight map, the global and local components of multiple exposure-adjusted images are fused.

[0076] In one feasible implementation, when the local enhancement module 603 enhances local image patches in the preprocessed image based on an enhancement control factor to generate a locally enhanced image, it converts the preprocessed image from the RGB color space to the CIELAB color space to calculate the mean and variance of the local image patches in the luminance channel in the CIELAB color space; determines the enhancement control factor based on the ratio of the global variance to the variance of the local image patches, and enhances the contrast of the local image patches based on the enhancement control factor to obtain the enhanced image; uses a guided filter to suppress noise in the enhanced image to obtain a denoised image; and converts the denoised image from the CIELAB color space to the RGB color space to obtain the locally enhanced image.

[0077] In one feasible implementation, the preprocessing module 601 preprocesses the foggy remote sensing image based on an atmospheric scattering model. When generating the preprocessed image, it is used to obtain atmospheric light parameters based on the foggy remote sensing image. Based on the atmospheric light parameters, the foggy remote sensing image is preprocessed using an atmospheric scattering model to generate the preprocessed image.

[0078] In one feasible implementation, when the global enhancement module 602 extracts the global and local components of each exposure adjustment image, it uses a guided filter to filter the brightness component in each exposure adjustment image to obtain the global component.

[0079] To demonstrate the beneficial effects of the above embodiments, this application provides a method for experimentally verifying the above-described remote sensing image dehazing method.

[0080] I. Experimental Setup Datasets: This embodiment uses the UAV, EuroSAT, SRI-WHU, and RICE datasets 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 is derived from Sentinel-2 satellite imagery and contains 13 spectral bands. It includes ten subsets of 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.

[0081] Baseline method: The AFAM method was compared with 10 image dehazing methods, including CGID, BCDP, GIFM, PMT, DNMGDT, DEA-Net, RIDCP, MSTN, SGID-PFF, and SSID.

[0082] Implementation Details: To ensure fair comparisons, the results reported by AFAM and all comparison methods were performed on the same device and with the same configuration settings. For baseline methods, this embodiment uses their source code and reports the reproduced performance. For deep learning-based methods, the ADAM optimizer with default parameters and an initial learning rate of 0.0001 was used. This implementation was performed in a PyCharm 2022 environment using Python 3.7 and CUDA 11.6. For traditional methods, MATLAB R2020a was used for implementation.

[0083] Evaluation Metrics: Four commonly used image quality 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 its 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 patches of an image, 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.

[0084] A. UAV dataset Qualitative comparison: The dehazing effects of different methods on synthetic hazy RS images tested on a drone dataset, such as... Figure 7 As shown, Figure 7 In the images, (a) is a raw image; (b) is CGID; (c) is BCDP; (d) is GIFM; (e) is PMT; (f) is DNMGDT; (g) is DEA-Net; (h) is RIDCP; (i) is MSTN; and (j) is Ours. The results show that this method, combined with GIFM, effectively improves the clarity of the UAV. However, some deep learning methods, such as DEA-Net, MSBDN, and FFA, are almost unable to completely eliminate this haze. The remaining traditional methods can only partially remove a small amount of haze. In terms of color correction, CAP consistently produces low-contrast results and cannot effectively correct color distortion. In contrast, this method performs well in both dehazing and color correction.

[0085] Quantitative comparison: Figure 8 The scores of different methods on the UAV dataset for AG, EI, PCQI, and IE are displayed. Our method achieved the highest or near-highest score in the quantitative evaluation of this dataset. While GIFM also yielded good results, its computational complexity is high, making real-time processing impractical. Overall, for synthetic blurred RS images, our method outperforms other comparative methods in both qualitative and quantitative evaluations.

[0086] B. EuroSAT Dataset Qualitative comparison: Figure 9 The image shows a visual comparison of different dehazing methods on the EuroSAT dataset, where (a) is a raw image; (b) is CGID; (c) is BCDP; (d) is GIFM; (e) is PMT; (f) is DNM-GDT; (g) is DEA-Net; (h) is RIDCP; (i) is MSTN; and (j) is Ours. Prior-based methods (CGID and BCDP) can only partially remove haze, leaving obvious gray veils and flattened color areas. While GIFM and PMT enhance global brightness, they easily introduce overexposure, weakening local contrast in vegetated areas and resulting in unrealistic tonal distributions. Deep learning-based methods, such as DNM-GDT, DEA-Net, and MSTN, recover structural details to some extent, but often exhibit slight color shifts or incomplete haze suppression at distances. RIDCP produces relatively clear results, but amplifies local artifacts, leading to inconsistent lighting across different plots. In contrast, AFAM in this embodiment produces a more natural and visually coherent reconstruction.

[0087] Quantitative comparison: such as Figure 8 As shown, among all evaluation metrics, including gradient-based, edge-based, and perception-based metrics, the proposed AFAM achieved the most favorable quantitative results on the EuroSAT dataset. Compared with CGID, BCDP, RIDCP, DEA-Net, and MSTN, AFAM shows significant improvement in structure-oriented measures, reflecting its stronger ability to recover fine-scale edge and texture continuity in complex agricultural and urban layouts.

[0088] C.SIRI-WHU dataset Qualitative comparison: Figure 10Visual comparisons of various dehazing methods in the SR-WHU dataset are provided, where (a) is a raw image; (b) is CGID; (c) is BCDP; (d) is GIFM; (e) is PMT; (f) is DNMGDT; (g) is DEA-Net; (h) is RIDCP; (i) is MSTN; and (j) is Ours. This dataset contains large-scale urban and rural scenes with strong illumination variations. CGID and BCDP fail to completely remove haze, producing flat contrast and gray tones, especially in densely built-up areas. While GIFM and PMT improve overall brightness, they tend to overexpose high-reflectivity areas and distort color consistency between roads and vegetation. Deep learning-based models such as DNMGDT and DEA-Net can recover some structural details but often leave residual haze in shadowed areas or introduce localized darkness. RIDCP and MSTN perform relatively well in maintaining structural integrity but still suffer from uneven brightness and suppression of fine textures. In contrast, AFAM in this embodiment produces the clearest and most visually natural results, with sharp building edges, vibrant vegetation colors, and balanced brightness throughout the scene.

[0089] Quantitative comparison: such as Figure 8 As shown, the proposed AFAM achieves the most competitive results across all quantitative metrics on the SIRI-WHU dataset. Compared to traditional prior-based deep learning models and recent deep learning models, AFAM demonstrates significant advantages in edge and gradient-related metrics, highlighting its ability to maintain geometric fidelity and structural clarity in large-scale remote sensing scenes. Perceptual and information-based metrics also show significant improvements, indicating that AFAM maintains consistent brightness and natural color contrast even in regions with complex lighting and texture transitions.

[0090] D. Enhanced Detail Analysis RS images with clear texture details are of great significance in various fields, including agriculture and industry. For example... Figure 11 As shown, the detail enhancement performance of this method and related methods is further evaluated, with certain areas magnified for better viewing. This embodiment selects a representative image from the SRI-WHU dataset for comparison. From a global perspective, this method demonstrates superior performance in color correction, contrast enhancement, and haze removal. Furthermore, in terms of local enhancement performance, such as… Figure 11 As shown in red in the image, this method significantly enhances the texture details and contrast of RS images compared to other methods.

[0091] E. Ablation test To demonstrate the beneficial effects of each core module in this method, this embodiment uses four objective evaluation metrics: AG, EI, PCQI, and IE, to conduct subsequent experimental analysis on the dataset: (a) DCP (AFAM without dark channel preprocessing). (b) MFEE (AFAM without multi-frequency exposure enhancement). (c) AFCM (AFAM without adaptive frequency compensation mapping). (d) AFA (AFAM without dominant feature aggregation). Results are shown in […]. Figure 13 .

[0092] Figure 12 Qualitative results of the AFAM method tested on three datasets are presented. The performance and strength of each module can be observed. (a) DCP: Preserves details of RS images, but dehazing performance needs improvement. (b) MFEE: Enhances local contrast, but the overall contrast enhancement effect is poor. (c) AFCM: It performs well in improving global contrast, but is insufficient in enhancing local contrast. (d) AFA: It cannot effectively fuse the advantages of global and local contrast, resulting in unsatisfactory fusion results. The full model not only successfully fuses the advantages of global and local contrast, but also effectively eliminates image blur, producing visually high-quality output results.

[0093] In this embodiment, a RS image dehazing method called AFAM is proposed. This method achieves effective dehazing by aggregating multi-frequency information through a dual-branch enhancement and fusion strategy. Specifically, the MFEE and AFCM modules are used to compensate for high-frequency texture degradation and enhance low-frequency details and local brightness. The two branches are further integrated by the AFA module to perform advantageous feature aggregation and reconstruct a high-quality haze-free image. Extensive experiments on three benchmark datasets show that AFAM consistently outperforms existing state-of-the-art methods.

[0094] Combination Figure 14 As shown, this application embodiment provides an electronic device 1400, including a processor 1401 and a memory 1402. Optionally, the device may further include a communication interface 1403 and a bus 1404. The processor 1401, communication interface 1403, and memory 1402 can communicate with each other via the bus 1404. The communication interface 1403 can be used for information transmission. The processor 1401 can call logical instructions in the memory 1402 to execute the remote sensing image dehazing method in the above embodiment.

[0095] Furthermore, the logic instructions in the aforementioned memory 1402 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0096] The memory 1402, 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 1401 executes functional applications and data processing by running the program instructions / modules stored in the memory 1402, thereby implementing the remote sensing image dehazing method in the above embodiments.

[0097] The memory 1402 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 1402 may include high-speed random access memory and may also include non-volatile memory.

[0098] This application provides a storage medium storing computer-executable instructions, which are configured to execute the remote sensing image dehazing method described in the above embodiments.

[0099] The aforementioned storage medium can be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0100] 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.

[0101] This application provides a computer program product, including a computer program, which, when executed by a processor, implements the access control method described above.

[0102] 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.

[0103] 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.

[0104] 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: Acquire foggy remote sensing images and preprocess the foggy remote sensing images based on an atmospheric scattering model to generate preprocessed images; Gamma correction is performed on the preprocessed image to generate multiple exposure-adjusted images, and a global enhancement image is generated based on the multiple exposure-adjusted images; The local image patches in the preprocessed image are enhanced based on the enhancement control factor to generate a locally enhanced image; By fusing the global enhanced image and the local enhanced image, a dehazed remote sensing image is obtained.

2. The method according to claim 1, characterized in that, The step of performing gamma correction on the preprocessed image to generate multiple exposure-adjusted images, and then generating a globally enhanced image based on the multiple exposure-adjusted images, includes: Gamma correction is performed on the preprocessed image based on multiple gamma values ​​to generate multiple exposure-adjusted images; Extract the global and local components of each of the exposure-adjusted images; The global and local components of multiple exposure-adjusted images are fused to generate a globally enhanced image.

3. The method according to claim 2, characterized in that, The fusion of global and local components from multiple exposure-adjusted images includes: Construct a global weight graph and local weight graphs; Based on the global weight map and the local weight map, the global and local components of multiple exposure-adjusted images are fused.

4. The method according to claim 1, characterized in that, The enhancement process based on enhancement control factors on local image patches in the preprocessed image to generate a locally enhanced image includes: The preprocessed image is converted from the RGB color space to the CIELAB color space to calculate the mean and variance of local image patches in the luminance channel in the CIELAB color space. The enhancement control factor is determined based on the ratio of the global variance to the variance of the local image patch, and the contrast of the local image patch is enhanced based on the enhancement control factor to obtain the enhanced image. A guided filter is used to suppress noise in the enhanced image, resulting in a denoised image. The denoised image is converted from the CIELAB color space to the RGB color space to obtain a locally enhanced image.

5. The method according to claim 1, characterized in that, The preprocessing of the foggy remote sensing image based on the atmospheric scattering model to generate a preprocessed image includes: Atmospheric light parameters are obtained from the foggy remote sensing image. The foggy remote sensing image is preprocessed using the atmospheric scattering model based on the atmospheric light parameters to generate the preprocessed image.

6. The method according to claim 2, characterized in that, The extraction of global and local components from each of the exposure-adjusted images includes: The global component is obtained by filtering the luminance component in each of the exposure-adjusted images using a guided filter.

7. A remote sensing image dehazing device, characterized in that, include: The preprocessing module is used to acquire foggy remote sensing images and preprocess the foggy remote sensing images based on an atmospheric scattering model to generate preprocessed images. A global enhancement module is used to perform gamma correction on the preprocessed image and generate multiple exposure adjustment images, so as to generate a global enhancement image based on the multiple exposure adjustment images; The local enhancement module is used to enhance local image patches in the preprocessed image based on enhancement control factors to generate a locally enhanced image. The fusion module is used to fuse the global enhanced image and the local enhanced image to obtain a dehazed remote sensing image.

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

Images