Remote sensing image defogging method and system based on multi-frequency dominant feature aggregation

Abstract

The invention discloses a remote sensing image defogging method and system based on multi-frequency dominant feature aggregation. The method comprises the following steps: performing multi-frequency exposure enhancement processing on an input foggy remote sensing image to obtain a global enhancement component and a local enhancement component; performing adaptive frequency compensation processing to generate a low frequency compensation enhanced image; constructing multi-scale representation and calculating a low-frequency component of each scale; determining high-frequency advantage features in the multi-scale space and fusing the high-frequency advantage features with the low-frequency components to form multi-scale fusion features; and completing the reconstruction of the fog-free remote sensing image based on the multi-scale fusion features. The system structurally comprises an image input module, a multi-frequency exposure enhancement module, an adaptive frequency compensation module, a multi-scale feature processing module, a dominant feature aggregation module and an image reconstruction module. According to the method, the brightness structure and the detail texture can be simultaneously enhanced and the residual haze can be inhibited in a complex remote sensing scene, so that a clear and natural fog-free remote sensing image with good structural consistency can be obtained.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method and system for dehazing remote sensing images based on the aggregation of multi-frequency advantageous features. Background Technology

[0002] Remote sensing images are of great value in fields such as aerospace observation, environmental monitoring, agricultural and forestry surveys, emergency rescue, and land analysis. Their imaging process is typically affected by factors such as long distance, large scale, high imaging altitude, and complex lighting conditions. During propagation in the atmosphere, water vapor, smoke, and pollutants in the air scatter and attenuate light, causing haze degradation problems in the acquired images, such as reduced brightness, insufficient contrast, and blurred details. This significantly affects the usability of remote sensing images and the accuracy of subsequent analysis tasks. Therefore, research on dehazing and quality enhancement of remote sensing images is of great significance.

[0003] Existing methods for dehazing remote sensing images mainly include physical model-based methods and data-driven methods. Physical model-based methods typically rely on atmospheric scattering models to recover haze-free images by estimating transmittance and atmospheric light. However, they are prone to color casts, artifacts, and inaccurate transmittance estimations in complex terrain features, locally overexposed or shadowed areas. Deep learning-based methods can learn more complex degradation features, but they are highly dependent on training data and struggle to adapt to imaging differences across regions, seasons, and sensors. In real-world remote sensing scenarios, high-frequency textures (such as building edges and road structures) and low-frequency brightness information (such as regional fog layers and uniform fog distribution) degrade simultaneously, and existing models often struggle to balance multi-scale structure restoration and local brightness compensation. Furthermore, large-scale remote sensing image processing faces challenges such as high efficiency requirements, significant equipment heterogeneity, and high image complexity, resulting in significant shortcomings in robustness, generalization, and processing accuracy of traditional methods.

[0004] Against this backdrop, the main problem that remote sensing dehazing technology urgently needs to solve is: how to effectively restore high-frequency texture structure and low-frequency brightness details in images when facing complex atmospheric imaging conditions and multi-category ground scene, and maintain the stability and universality of the method while improving image clarity and structural consistency, so as to improve the usability of remote sensing images under multi-scene and multi-temporal conditions. Summary of the Invention

[0005] To address the problems of existing technologies, embodiments of the present invention provide a method and system for dehazing remote sensing images based on multi-frequency dominant feature aggregation. The technical solution is as follows: On the one hand, a remote sensing image dehazing method based on multi-frequency dominant feature aggregation is provided, including the following steps: (1) Perform multi-frequency exposure enhancement processing on the input foggy remote sensing image, generate a multi-exposure image sequence using multiple different exposure adjustment parameters, and extract the global enhancement component and local enhancement component based on the multi-exposure image sequence respectively; (2) Perform adaptive frequency compensation processing on the brightness channel of the foggy remote sensing image, obtain brightness statistics according to the preset local block division method, and adaptively enhance the brightness difference based on the brightness statistics, while setting enhancement amplitude limit conditions to generate a low-frequency compensated enhanced image. (3) Construct a multi-scale representation based on the global enhancement component and the low-frequency compensation enhancement image, obtain the corresponding low-frequency component at each scale and determine the high-frequency dominant features; (4) The low-frequency component and the high-frequency dominant feature are fused at each scale to obtain multi-scale fused features; (5) Reconstruct the image based on the multi-scale fusion features to generate a fog-free remote sensing image.

[0006] Further, the multi-frequency exposure enhancement process in step (1) includes: Multiple sets of different exposure adjustment parameters are used to adjust the exposure of the foggy remote sensing image. By changing the brightness gain, gamma mapping intensity or exposure offset, a multi-exposure image sequence is generated that simultaneously covers three exposure levels: dark, moderate, and bright, in order to improve the ability of the image to represent structural information in different brightness ranges.

[0007] Furthermore, the acquisition methods for the global enhancement component and the local enhancement component in step (1) include: The multi-exposure image sequence generated in step (1) is subjected to structural component extraction processing to obtain a global enhancement component that reflects the overall brightness change trend. Each exposure image is then differentially processed with the corresponding structural component to obtain a local enhancement component that characterizes local detail changes.

[0008] Furthermore, the fusion method of the global enhancement component and the local enhancement component in step (1) includes: Exposure weights are constructed based on the exposure characteristics of each exposure image. The global enhancement component and the local enhancement component obtained in step (1) are weighted respectively, and the weighted global enhancement component and the weighted local enhancement component are fused to form the multi-frequency exposure enhancement result in step (1).

[0009] Further, the adaptive frequency compensation process in step (2) includes: The brightness channel of the foggy remote sensing image is divided into blocks according to a preset local block division method. The brightness difference of each local block is adaptively enhanced according to the average brightness and brightness change range of each local block. An enhancement range limit is set to avoid excessive brightness enhancement in local areas, thereby obtaining the low-frequency compensation enhanced image in step (2).

[0010] Furthermore, the adaptive frequency compensation process in step (2) further includes: The brightness enhancement result obtained in step (2) is subjected to guided filtering. The noise generated during the brightness enhancement process is suppressed by local linear modeling, and the smooth transition of the brightness structure is maintained, so as to improve the visual stability of the low-frequency compensation enhancement image in step (2).

[0011] Furthermore, the multi-scale feature processing in steps (3) and (4) includes: In the multi-scale representation constructed in step (3), low-frequency components reflecting the overall brightness structure are extracted for each scale, and high-frequency advantageous features that can highlight the detail area are determined after comparing the local change characteristics of the global enhancement component and the low-frequency compensation enhancement image. In step (4), the low-frequency components of each scale are fused with the corresponding high-frequency advantageous features to form the multi-scale fusion features in step (4).

[0012] On the other hand, a remote sensing image dehazing system based on multi-frequency dominant feature aggregation is provided, including: (1) Multi-frequency exposure enhancement module, used to perform multi-frequency exposure enhancement on the input foggy remote sensing image, generate a multi-exposure image sequence, and extract global enhancement component and local enhancement component from the multi-exposure image sequence; (2) An adaptive frequency compensation module is used to perform adaptive frequency compensation on the brightness channel of the foggy remote sensing image, enhance the brightness difference based on local brightness statistics, and generate a low-frequency compensated enhanced image according to the enhancement amplitude limit. (3) Multi-scale representation construction module, used to construct multi-scale representation based on the global enhancement component and the low-frequency compensation enhancement image, and obtain the low-frequency component and high-frequency advantage features of each scale; (4) Multi-scale fusion module, used to fuse low-frequency components of each scale with corresponding high-frequency advantageous features to generate multi-scale fusion features; (5) Image reconstruction module, used to reconstruct fog-free remote sensing images based on the multi-scale fusion features.

[0013] On the other hand, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, is used to implement the method.

[0014] On the other hand, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the computer program is executed by the processor, it is used to implement the method.

[0015] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows: This invention provides a remote sensing image dehazing method based on multi-frequency advantageous feature aggregation. By introducing multi-frequency exposure enhancement processing and adaptive frequency compensation processing in the enhancement stage, the input hazy remote sensing image obtains global enhancement components, local enhancement components and low-frequency compensation enhanced images respectively, achieving complementary enhancement in terms of brightness structure, detail texture and regional contrast.

[0016] Based on this, the present invention constructs a multi-scale representation and uses a high-frequency dominant feature selection mechanism at each scale to extract high-frequency features from the global enhancement component and the local enhancement component. At the same time, the low-frequency compensated enhancement image is fused with the low-frequency component within the scale, so that high-frequency details are preserved and enhanced, while low-frequency brightness structure is effectively compensated, thereby improving the overall structural consistency and visual clarity of the image.

[0017] The multi-scale fusion features, achieved by aggregating advantageous features, enable the joint recovery of high- and low-frequency information in the final reconstruction. This results in fog-free remote sensing images that are significantly superior to traditional methods in terms of texture sharpness, detail fidelity, brightness balance, and natural color retention, exhibiting strong scene adaptability and robustness. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a remote sensing image dehazing method based on multi-frequency advantageous feature aggregation according to Embodiment 1 of the present invention.

[0020] Figure 2 This is a schematic diagram of a remote sensing image dehazing system based on multi-frequency advantageous feature aggregation according to Embodiment 2 of the present invention.

[0021] Figure 3 This is a sample image of the UAV dataset used in the experimental section of this invention.

[0022] Figure 4 This is a sample image of the EuroSAT dataset used in the experimental section of this invention.

[0023] Figure 5 This is a sample image of the SIRI-WHU dataset used in the experimental section of this invention.

[0024] Figure 6 This is a diagram showing the ablation experiment results of the experimental part of this invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0026] Example 1 This embodiment provides a remote sensing image dehazing method based on multi-frequency advantageous feature aggregation, which is suitable for dehazing foggy remote sensing images. Through multi-frequency exposure enhancement processing, adaptive frequency compensation processing, multi-scale feature processing, and image reconstruction, fog-free remote sensing images are acquired.

[0027] In this embodiment, the method takes a foggy remote sensing image as input and outputs a fog-free remote sensing image. See also Figure 1 The overall process of the method includes steps (1) to (5), as follows.

[0028] The method in this embodiment generally includes two main stages: an enhancement stage and an aggregation stage. In the enhancement stage, multi-frequency exposure enhancement processing and adaptive frequency compensation processing are performed on the foggy remote sensing image to obtain global enhancement components, local enhancement components, and low-frequency compensated enhanced images, respectively. In the aggregation stage, the above enhancement results are aggregated with advantageous features under multi-scale representation, and high- and low-frequency complementary information is fused to finally complete the reconstruction of the fog-free remote sensing image.

[0029] Step (1): Multi-frequency exposure enhancement processing In step (1), the input foggy remote sensing image is subjected to multi-frequency exposure enhancement processing. Multiple different exposure adjustment parameters are used to generate a multi-exposure image sequence, and global enhancement components and local enhancement components are extracted based on the multi-exposure image sequence.

[0030] 1. Preprocessing and Atmospheric Light Estimation of Foggy Remote Sensing Images In this embodiment, the input foggy remote sensing image is first preprocessed. Key parameters for estimating atmospheric light are extracted from the original image based on the dark channel prior, and the foggy remote sensing image is corrected using an atmospheric scattering model to obtain a preprocessed image.

[0031] The formula is as follows: I(x) = J(x)t(x) + A(1−t(x)) Where I(x) represents the observed image, J(x) represents the actual scene radiation image, t(x) represents the transmittance, and A represents the global atmospheric light.

[0032] The above preprocessing reduces atmospheric light interference in foggy areas, providing a more stable base input for subsequent multi-exposure enhancement.

[0033] 2. Generation of multi-exposure image sequences Based on the preprocessed image, multiple sets of exposure adjustment parameters are used to adjust the exposure of the preprocessed image. In this embodiment, the exposure adjustment parameters include brightness gain parameters, gamma mapping intensity parameters, and exposure offset parameters. By changing the gamma mapping intensity and brightness gain, the preprocessed image is used to form a multi-exposure image sequence at three exposure levels: dark, moderate, and bright.

[0034] The formula is as follows:

[0035] in, Indicates the pixel position of the preprocessed image. The intensity value at that location; Indicates pixel position; Indicates the brightness gain parameter; This represents the gamma mapping intensity parameter; This represents a mapping transformation.

[0036] After the above processing, a multi-exposure image sequence covering three exposure levels—dark, moderate, and bright—is obtained. This multi-exposure image sequence is used for subsequent extraction of global and local enhancement components.

[0037] 3. Acquisition of Global Augmentation Components For each exposure image in a multi-exposure image sequence, structural component extraction is performed on the brightness channel. The structural component extraction employs a guided filtering method, constructing a linear model between the input image and the guided image within a local window to obtain structural components that reflect the overall brightness variation trend.

[0038] The formula is as follows:

[0039] in, Indicates the first Structural components of an exposed image; Indicates the guided filter operator; Indicates the local window radius; Represents the regularization parameter; Indicates the first The input of the exposure image on the brightness channel.

[0040] After obtaining multiple sets of structural components, the multi-exposure structural components are analyzed, and a weighting function based on exposure characteristics is constructed according to the deviation between grayscale intensity and ideal exposure level. This weighting function is used to characterize the contribution of different exposure images in various brightness regions.

[0041] The formula is as follows:

[0042] in, This represents the global enhancement component after multi-exposure fusion; This indicates the structural components of each exposed image. The fusion result after weighting; This represents the brightness / exposure-related components of each exposed image (by...). The fusion result after weighting (characterization) is used to represent the contribution of different exposure images to different brightness regions.

[0043] The structural components corresponding to each exposed image are weighted and fused using the aforementioned exposure weights to obtain a global enhancement component after multi-exposure fusion. This global enhancement component is used to describe the overall brightness structural features. 4. Acquisition of Local Enhancement Components (Based on the obtained structural components, differential processing is performed on each exposed image and its corresponding structural components to obtain local difference components. Local difference components reflect the local detail changes relative to the overall brightness structure under the corresponding exposure conditions, and are used to describe texture, edges, and small-scale brightness variations.)

[0044] The formula is as follows:

[0045] in, Indicates the first Local difference components (local enhancement components) of an exposed image. Indicates the first A single exposure image (brightness / input image); Indicates the first The structural components (global brightness trend components) corresponding to the exposure image. This indicates pixel-by-pixel subtraction, used to remove overall brightness trends while preserving detail variations.

[0046] The above differential operation eliminates the influence of the overall brightness trend while preserving high-frequency information of prominent details and local contrast changes. After obtaining multiple sets of local differential components, based on the aforementioned exposure weights, the local differential components under different exposure conditions are weighted to obtain local enhancement components.

[0047] At the image representation level, local enhancement components can be organized into local enhancement images, which, together with the global enhancement images corresponding to the global enhancement components, serve as inputs for subsequent aggregation of advantageous features.

[0048] 5. Formation of multi-frequency exposure enhancement results In this embodiment, the structural component and the local differential component are weighted and fused by exposure weight to obtain the joint output of the global enhancement component and the local enhancement component, which can be regarded as the multi-frequency exposure enhancement result in step (1).

[0049] The formula is as follows:

[0050] in, This represents the final output image after multi-exposure enhancement; This represents the global enhancement component obtained through fusion; Indicates the first Local enhancement components of a multi-exposure image; This indicates pixel-by-pixel addition, used to fuse global enhancements and local detail enhancements to obtain the final result.

[0051] After the above multi-frequency exposure enhancement processing is completed, the global enhancement component and the local enhancement component obtained in step (1) will be used as inputs for multi-scale feature processing and advantageous feature aggregation, which is the basis of the "fusion method of the global enhancement component and the local enhancement component" in claim 4.

[0052] Step (2): Adaptive Frequency Compensation Processing (AFCM) In step (2), adaptive frequency compensation processing is performed on the brightness channel of the foggy remote sensing image. Brightness statistics are obtained according to the preset local block division method, and the brightness difference is adaptively enhanced based on the brightness statistics. At the same time, enhancement amplitude limit conditions are set to generate a low-frequency compensated enhanced image.

[0053] 1. Local block partitioning and acquisition of brightness statistics In this embodiment, the foggy remote sensing image is converted from the RGB color space to the CIELAB color space, the luminance channel is extracted, and the luminance channel is divided into multiple local blocks according to a preset local block partitioning method. For each local block, the mean luminance and luminance variance of the local block are calculated. The mean luminance is used to describe the overall luminance level of the local block, and the luminance variance is used to describe the luminance variation range and texture complexity of the local block.

[0054] The formula is as follows:

[0055] in, Indicates position The brightness statistics / compensation base value of the corresponding local block; This represents the average brightness of the local block. This represents the local block brightness variance; Indicates the global brightness variance; This indicates taking the minimum value, and is used to truncate and restrict the compensation term.

[0056] The formula is as follows:

[0057] in, Indicates position The brightness value at that location; Indicates position Brightness compensation bias at the location; This indicates pixel-by-pixel addition, used for brightness compensation adjustments.

[0058] 2. Adaptive brightness difference enhancement Based on the mean and variance of local block brightness, a local enhancement factor is adaptively constructed. Linear enhancement is performed on the brightness difference within each local block, so that blocks with severe fogging and reduced contrast receive greater compensation during the enhancement process, while blocks with relatively normal brightness and contrast only receive limited enhancement.

[0059] The formula is as follows:

[0060] in, Indicates position The brightness value after adaptive brightness difference enhancement; This represents the average brightness of the local block. Indicates position The brightness value at that location (belonging to a local block) ); Indicates position Local enhancement factor at the location; This represents the brightness difference relative to the local mean.

[0061] To avoid excessive enhancement of local blocks, the adaptive frequency compensation process sets an upper limit on the local enhancement factor by comparing the local block brightness variance with the global brightness variance, thereby limiting the enhancement magnitude.

[0062] The formula is as follows:

[0063] in, Indicates position Local enhancement factor at the location; This represents the local block brightness variance; This represents the global brightness variance; this ratio is used to adaptively adjust the enhancement intensity based on the magnitude of local texture / brightness changes.

[0064] After the above adaptive brightness difference enhancement and limiting processing, the brightness channel is compensated for in terms of low-frequency contrast.

[0065] 3. Guided filtering for noise reduction and low-frequency compensation to enhance image generation To suppress noise introduced during adaptive enhancement while maintaining a smooth transition in the brightness structure, the adaptive frequency compensation process further applies guided filtering to the enhanced brightness result. Guided filtering constructs a linear model within a local region, establishing a linear relationship between the enhanced brightness and the reference brightness, thus achieving noise suppression and edge preservation.

[0066] The formula is as follows:

[0067] in, Indicates the output of the guided filter at the pixel The value at; Indicates the guide image in pixels The value at; , Display window The coefficients of the linear model within; Indicates A local window centered on the user; This indicates that the linear relationship holds true for all pixels within the window.

[0068] The brightness result after guided filtering is used as a low-frequency compensated enhanced image.

[0069] Steps (3) and (4): Multi-scale feature processing and advantageous feature aggregation (AFA) In steps (3) and (4), multi-scale feature processing is performed on the global enhancement component, local enhancement component and low-frequency compensation enhancement image. By constructing multi-scale representation, extracting low-frequency component and high-frequency dominant feature, and fusing low-frequency component and high-frequency dominant feature at each scale, multi-scale fusion feature is obtained.

[0070] 1. Multi-scale representation construction and low-frequency component acquisition In this embodiment, a multi-scale representation is constructed based on the global enhancement component and the low-frequency compensation enhanced image. Specifically, the Laplacian pyramid method can be used to decompose the global enhancement component and the low-frequency compensation enhanced image into low-frequency components at multiple scales.

[0071] At each scale, low-frequency components based on global enhancement components and low-frequency components based on low-frequency compensation enhancement images are obtained separately, and the two are combined to form the low-frequency component at that scale.

[0072] The formula is as follows:

[0073] in, Indicates position Low-frequency components at the location; Indicates the first The low-frequency component is located at... The value at; This represents the corresponding weighting coefficient; This indicates the number of low-frequency components involved in the combination.

[0074] This low-frequency component is used to represent the overall brightness structure and basic contrast at this scale, which corresponds to the design in the technical disclosure document that "Mt is the average of the two low-frequency components".

[0075] 2. Determination of high-frequency dominant characteristics Within each scale, the high-frequency structures obtained from global and local enhanced images are compared. By analyzing the deviation of their ratios from the corresponding low-frequency components, the dominant high-frequency features for highlighting details at that scale are determined. Specifically, within the same scale, the side with a larger high-frequency response amplitude and clearer local structure is selected as the dominant high-frequency feature. The formula is as follows:

[0076] in, Indicates position High-frequency dominant features / high-frequency components at the location; This indicates the enhancement results (or high-frequency responses) at the same scale at location. The value at; Indicates the location of low-frequency components at the same scale. The value at that location.

[0077] A small positive parameter can be introduced in the ratio calculation to avoid the denominator being zero, thus ensuring the numerical stability of the high-frequency dominant feature calculation process. The high-frequency dominant feature corresponds to the specific implementation of the "high-frequency dominant feature" in claims 1 and 7.

[0078] 3. Formation of multi-scale fusion features and cross-scale aggregation Within each scale, the low-frequency component of that scale is combined with the corresponding high-frequency dominant feature according to a preset fusion method to generate the fused feature of that scale.

[0079] The formula is as follows:

[0080] in, Representing scale The low-frequency fusion results are as follows; Indicates position Fusion coefficient / gain coefficient at the location; Representing scale The low-frequency components below are in position The value at that location.

[0081] Subsequently, the fusion features at multiple scales are aggregated from bottom to top to generate multi-scale fusion features for image reconstruction.

[0082] The formula is as follows:

[0083] in, Indicates the location The sum of the normalized weights for each scale / component is: Indicates the first Each component in the scale Below, position Weight at each location; This indicates a non-negative weight constraint.

[0084] Step (5): Reconstruction of fog-free remote sensing images In step (5), image reconstruction is performed based on multi-scale fusion features to generate fog-free remote sensing images.

[0085] In this embodiment, by performing scale-by-scale reconstruction on the fused features at each scale, the multi-scale fused features are mapped back to the original resolution space to obtain a fog-free remote sensing image. This reconstruction process can be achieved by weighted summation or scale-by-scale inverse transformation of the multi-scale fused features and low-frequency components. The formula is as follows:

[0086] in, This indicates the location of the reconstructed fog-free image. Pixel value at; This represents the low-frequency component after multi-scale fusion; This indicates the high-frequency advantageous features after multi-scale fusion; This indicates pixel-by-pixel addition, used to fuse basic brightness structure and detailed information.

[0087] Through the above steps, the final fog-free remote sensing image is superior to the input foggy remote sensing image in terms of overall brightness, local contrast, and detail structure.

[0088] In summary, the remote sensing image dehazing method based on multi-frequency dominant feature aggregation described in this embodiment introduces multi-frequency exposure enhancement processing and adaptive frequency compensation processing in the enhancement stage. On the one hand, it utilizes global enhancement components and local enhancement components to synergistically enhance the overall brightness structure and local texture details in the fogged scene; on the other hand, it enhances the image recovery of brightness and contrast attenuation caused by atmospheric scattering through low-frequency compensation. In the aggregation stage, low-frequency components are constructed through multi-scale representation, and high-frequency dominant features are selected at each scale. The multi-scale fusion features are used for the reconstruction of fog-free remote sensing images, resulting in output images with high consistency and visual readability in terms of overall brightness, local contrast, and preservation of detail structure. The steps given in this embodiment have clear data flow relationships, and the meaning of parameters and processing flow have been explained in the foregoing, which is sufficient to enable those skilled in the art to implement this method.

[0089] Example 2 This embodiment provides a remote sensing image dehazing system based on multi-frequency dominant feature aggregation. This system is suitable for processing hazy remote sensing images and outputting dehazed remote sensing images. See also... Figure 2 The overall system structure includes an input module, a multi-frequency exposure enhancement module, an adaptive frequency compensation module, a multi-scale feature processing module, a superior feature aggregation module, and an image reconstruction module. These modules are connected in sequence to form a complete image dehazing process.

[0090] The system first receives foggy remote sensing images through the input module and then transmits them to the multi-frequency exposure enhancement module. The multi-frequency exposure enhancement module performs image preprocessing, exposure adjustment, and generation of multi-exposure image sequences on the input image. This module also includes structural component extraction and local difference component extraction functions. Through weighted fusion of structural and local difference components, global and local enhancement components are obtained respectively, used to characterize the overall brightness structure and local detail changes of the image.

[0091] The adaptive frequency compensation module in the system processes the luminance channel of the original image, compensating for low-frequency contrast loss caused by fog effects through local block partitioning, luminance statistics calculation, and adaptive enhancement mapping. This module further generates a low-frequency compensated enhanced image by limiting the enhancement amplitude and using guided filtering for noise reduction, making the luminance structure more stable and providing a reliable low-frequency foundation for subsequent multi-scale processing.

[0092] In the multi-scale feature processing module, the system constructs a multi-scale representation based on global enhancement components, local enhancement components, and low-frequency compensation enhancement images. Each component is decomposed at multiple scales to extract low-frequency components, while high-frequency details are effectively separated during the decomposition process, enabling the system to accurately capture structural changes in the image at different scales.

[0093] The dominant feature aggregation module analyzes high-frequency structural features across multiple scales. By comparing the normalized deviation of high-frequency structures relative to low-frequency components, it identifies the more expressive high-frequency dominant features at that scale. Subsequently, this module fuses the high-frequency dominant features with the corresponding low-frequency components to obtain multi-scale fused features that incorporate comprehensive information on structure, texture, and contrast.

[0094] The image reconstruction module receives multi-scale fused features and maps these features back to the original image resolution space through scale-by-scale reconstruction or inverse transformation, forming the final fog-free remote sensing image. The reconstructed image outperforms the input image in terms of contrast, detail sharpness, and structure preservation, which is beneficial for subsequent interpretation and analysis of remote sensing images.

[0095] The remote sensing image dehazing system described in this embodiment adopts a modular design, with clearly defined functions for each module and a clear data flow. Through a series of processing techniques, including multi-frequency exposure enhancement, adaptive low-frequency compensation, multi-scale analysis, and fusion of advantageous features, it collaboratively restores the brightness, contrast, and detail of foggy remote sensing images, improving the dehazing effect while maintaining image structural consistency and detail integrity. This system can be deployed on remote sensing image processing platforms, image analysis software systems, or as part of embedded devices for airborne or ground-based processing scenarios, demonstrating high practical application value.

[0096] To verify the effectiveness of the remote sensing image dehazing method based on multi-frequency dominant feature aggregation proposed in this invention, experiments were conducted using three publicly available remote sensing image datasets, and the method was compared and evaluated with several representative dehazing methods. The experimental process included dataset setup, selection of comparison methods, implementation details, and qualitative and quantitative result analysis.

[0097] A. Experimental Setup Dataset The effectiveness of the proposed method was validated on three publicly available remote sensing dehazing datasets: UAV Dataset (Zheng and Zhang 2023): A partially publicly available dataset of remote sensing haze images used to evaluate the dehazing performance of different methods.

[0098] The EuroSAT dataset (Helber et al. 2019) is constructed from Sentinel-2 satellite imagery. It contains 13 spectral bands, covering 10 different scenes, and a total of 27,000 labeled and georeferenced images. Each scene category contains 2,000–3,000 images of size 64×64.

[0099] SIRI-WHU dataset (Zhao et al. 2015): A large-scale, high-resolution remote sensing image set used to test dehazing generalization capabilities in complex scenes.

[0100] Comparison Methods The proposed AFAM method was compared with eight defogging methods, including: CGID (Feng et al. 2024b), DNMGDT (Su et al. 2025), PMT (Feng et al. 2024a), DEA-Net (Chen, He, and Lu 2024), RIDCP (Wu et al. 2023), MSTN (Zhao, Zhang, and Cui 2022), SGID-PFF (Bai et al. 2022), SSID (Chen et al. 2025).

[0101] Implementation details To ensure a fair comparison, all methods were run on the same devices and configurations. Traditional methods were implemented using MATLAB R2020a, while deep learning methods were run using Python 3.7 + CUDA 11.6 in a PyCharm 2022 environment. The optimizer was Adam, with an initial learning rate of 0.0001.

[0102] Evaluation indicators Four commonly used image quality metrics were used: AG (Average Gradient) (David 2004): measures image sharpness; EI (Edge Intensity) (Azmi et al. 2019): reflects edge contrast; PSNR (Peak Signal-to-Noise Ratio); and SSIM (Structural Similarity Index) (Wang et al. 2004).

[0103] B. Experimental Results Qualitative comparison On the UAV dataset ( Figure 3 DEA-Net and CGID exhibit significant color distortion and loss of detail, while AFAM effectively recovers natural tones and fine textures. On the EuroSAT dataset (…), Figure 4 AFAM significantly suppressed residual haze and improved the contrast of terrain features. In complex urban scenes on the SIRI-WHU dataset ( Figure 5AFAM performs best in maintaining building boundaries and road layout. Overall, AFAM demonstrates stronger robustness and visual consistency across various scenarios.

[0104] Quantitative comparison Table 1 presents the quantitative metrics on the three major datasets. AFAM achieved the best performance across all evaluation metrics. For example, on the UAV dataset, AFAM's AG=5.716, EI=59.662, PSNR=16.372, and SSIM=0.8024 comprehensively outperformed the comparison methods. On the EuroSAT dataset, AFAM achieved the highest values ​​in AG, EI, and SSIM, indicating its effective preservation of structural details across multiple scenarios. On the SIRI-WHU dataset, AFAM achieved a PSNR of 22.421 dB, significantly exceeding other methods, and also achieved the best results in SSIM and AG.

[0105] Table 1 Comparative Experiments

[0106] C. Ablation test To verify the contribution of each module, ablation analysis was performed on the three core modules of AFAM: w / o MFEE (removal of multi-frequency exposure enhancement): high-frequency texture recovery capability decreased, leading to a significant reduction in PSNR and SSIM. w / o AFCM (removal of adaptive frequency compensation): insufficient low-frequency detail recovery, resulting in reduced EI and AG. w / o AFA (removal of advantageous feature aggregation): insufficient high- and low-frequency fusion, leading to an overall performance decline. Table 2 shows that the complete AFAM outperforms the variant with any module removed in all metrics, verifying the complementarity of the modules. Figure 6 It demonstrates the contribution of different components to visual enhancement.

[0107] Table 2 Ablation Experiment

[0108] D. Time complexity Table 3 shows the average runtime of different methods on 256×256 and 512×512 image patches. The results show that AFAM outperforms most state-of-the-art methods while maintaining optimal performance. For example, at 256×256 resolution, AFAM's average runtime is only 0.029s, much faster than SGID-PFF (0.413s) and DEA-Net (2.053s).

[0109] Table 3 Time Complexity

[0110] This invention proposes a remote sensing image dehazing method based on multi-frequency dominant feature aggregation. Through modular design including multi-frequency exposure enhancement, adaptive frequency compensation mapping, and dominant feature aggregation, it achieves comprehensive restoration of high-frequency texture information and low-frequency brightness structure. At the high-frequency level, multi-frequency exposure enhancement compensates for detail loss caused by fogging; at the low-frequency level, adaptive frequency compensation mapping effectively enhances local brightness and overall contrast; at the fusion level, the dominant feature aggregation mechanism selects features with superior expressive power for fusion based on the feature response intensity at different scales, thereby generating clear, natural, and structurally coherent fog-free remote sensing images. Extensive experimental results show that this invention outperforms existing technologies in multiple evaluation metrics such as AG, EI, PSNR, and SSIM. Component ablation experiments further verify the rationality of the module design, and operational efficiency analysis also proves that this invention maintains high-quality dehazing results while possessing high computational efficiency. In summary, this invention has good robustness, applicability, and engineering application value.

[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dehazing remote sensing images based on multi-frequency dominant feature aggregation, characterized in that, Includes the following steps: (1) Perform multi-frequency exposure enhancement processing on the input foggy remote sensing image, generate a multi-exposure image sequence using multiple different exposure adjustment parameters, and extract the global enhancement component and local enhancement component based on the multi-exposure image sequence respectively; (2) Perform adaptive frequency compensation processing on the brightness channel of the foggy remote sensing image, obtain brightness statistics according to the preset local block division method, and adaptively enhance the brightness difference based on the brightness statistics, while setting enhancement amplitude limit conditions to generate a low-frequency compensated enhanced image. (3) Construct a multi-scale representation based on the global enhancement component and the low-frequency compensation enhancement image, obtain the corresponding low-frequency component at each scale and determine the high-frequency dominant features; (4) The low-frequency component and the high-frequency dominant feature are fused at each scale to obtain multi-scale fused features; (5) Reconstruct the image based on the multi-scale fusion features to generate a fog-free remote sensing image.

2. The method according to claim 1, characterized in that, The multi-frequency exposure enhancement process in step (1) includes: Multiple sets of different exposure adjustment parameters are used to adjust the exposure of the foggy remote sensing image. By changing the brightness gain, gamma mapping intensity or exposure offset, a multi-exposure image sequence is generated that simultaneously covers three exposure levels: dark, moderate, and bright, in order to improve the ability of the image to represent structural information in different brightness ranges.

3. The method according to claim 1, characterized in that, The methods for obtaining the global enhancement component and the local enhancement component in step (1) include: The multi-exposure image sequence generated in step (1) is subjected to structural component extraction processing to obtain a global enhancement component that reflects the overall brightness change trend. Each exposure image is then differentially processed with the corresponding structural component to obtain a local enhancement component that characterizes local detail changes.

4. The method according to claim 1, characterized in that, The fusion method of the global enhancement component and the local enhancement component in step (1) includes: Exposure weights are constructed based on the exposure characteristics of each exposure image. The global enhancement component and the local enhancement component obtained in step (1) are weighted respectively, and the weighted global enhancement component and the weighted local enhancement component are fused to form the multi-frequency exposure enhancement result in step (1).

5. The method according to claim 1, characterized in that, The adaptive frequency compensation process in step (2) includes: The brightness channel of the foggy remote sensing image is divided into blocks according to a preset local block division method. The brightness difference of each local block is adaptively enhanced according to the average brightness and brightness change range of each local block. An enhancement range limit is set to avoid excessive brightness enhancement in local areas, thereby obtaining the low-frequency compensation enhanced image in step (2).

6. The method according to claim 1, characterized in that, The adaptive frequency compensation process in step (2) further includes: The brightness enhancement result obtained in step (2) is subjected to guided filtering. The noise generated during the brightness enhancement process is suppressed by local linear modeling, and the smooth transition of the brightness structure is maintained, so as to improve the visual stability of the low-frequency compensation enhancement image in step (2).

7. The method according to claim 1, characterized in that, The multi-scale feature processing in steps (3) and (4) includes: In the multi-scale representation constructed in step (3), low-frequency components reflecting the overall brightness structure are extracted for each scale, and high-frequency advantageous features that can highlight the detail area are determined after comparing the local change characteristics of the global enhancement component and the low-frequency compensation enhancement image. In step (4), the low-frequency components of each scale are fused with the corresponding high-frequency advantageous features to form the multi-scale fusion features in step (4).

8. A remote sensing image dehazing system based on multi-frequency dominant feature aggregation, characterized in that, include: (1) Multi-frequency exposure enhancement module, used to perform multi-frequency exposure enhancement on the input foggy remote sensing image, generate a multi-exposure image sequence, and extract global enhancement component and local enhancement component from the multi-exposure image sequence; (2) An adaptive frequency compensation module is used to perform adaptive frequency compensation on the brightness channel of the foggy remote sensing image, enhance the brightness difference based on local brightness statistics, and generate a low-frequency compensated enhanced image according to the enhancement amplitude limit. (3) Multi-scale representation construction module, used to construct multi-scale representation based on the global enhancement component and the low-frequency compensation enhancement image, and obtain the low-frequency component and high-frequency advantage features of each scale; (4) Multi-scale fusion module, used to fuse low-frequency components of each scale with corresponding high-frequency advantageous features to generate multi-scale fusion features; (5) Image reconstruction module, used to reconstruct fog-free remote sensing images based on the multi-scale fusion features.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the method of any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it is used to implement the method according to any one of claims 1 to 7.

Images