Remote sensing image defogging method and device based on wavelet multi-scale decomposition

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image defogging method and device based on wavelet multi-scale decomposition, and the method comprises the steps: dividing an original foggy remote sensing image into a low-frequency sub-band and a high-frequency sub-band through multi-scale wavelet frequency decomposition, and separating haze interference and image details; secondly, details of the high-frequency sub-bands are enhanced through a nonlinear dynamic enhancement function, multi-scale guiding filtering is conducted on the low-frequency sub-bands through the enhanced high-frequency sub-bands, and fog components in the image are restrained; further integrating feature information under different scales through residual cavity convolution and an SE module to reinforce the perception ability of a global structure and local details, and obtaining fusion features; and finally, decoding and reconstructing the fused features, and recovering to obtain a high-resolution defogged remote sensing image. According to the method, the definition, the structural integrity and the application adaptability of the remote sensing image after recovery under mild to severe haze conditions are effectively improved, the operation consumption is reduced, and the algorithm execution speed is improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and apparatus for dehazing remote sensing images based on wavelet multi-scale decomposition. Background Technology

[0002] Remote sensing images are often affected by atmospheric interference such as haze, fog, and thin clouds, which greatly hinders ground observation and subsequent target understanding and analysis tasks. Remote sensing image dehazing plays a crucial role in improving data availability for applications such as environmental monitoring, disaster response, and military reconnaissance.

[0003] Remote sensing image dehazing methods can generally be divided into traditional dehazing methods and deep learning-based dehazing methods.

[0004] Traditional dehazing methods mostly rely on image enhancement and physical models, including histogram equalization (HE) and dark channel prior (DCP). Histogram equalization, a traditional image enhancement technique, aims to broaden the grayscale value distribution range to enhance overall image contrast. This technique is highly effective for images with concentrated grayscale values ​​or low contrast, making it well-suited for addressing insufficient contrast between background and foreground. However, it still has limitations in dehazing remote sensing images. The dark channel prior method utilizes the statistical characteristics of the dark channel in hazy images to estimate the transmittance and atmospheric light value of hazy images. By analyzing the intensity distribution of the dark channel in hazy images, it estimates the contribution of atmospheric light and then calculates the transmittance to reconstruct the hazy image. When processing images with sky, dark channel priors are prone to color distortion because sky areas often have relatively high brightness values. Therefore, when estimating transmittance, dark channel priors may mistakenly treat sky areas as atmospheric light, resulting in color shifts. In areas with abrupt changes in depth of field, such as where mountains meet the sky, dark channel priors may also create halos. This is because the method does not adequately consider the differences in local features when estimating transmittance, leading to the loss of detailed information in boundary areas.

[0005] Traditional remote sensing image dehazing methods mostly rely on physical models or image enhancement techniques. While these methods improve image quality to some extent, they have significant shortcomings in areas such as detail preservation, color correction, and adaptability to complex and changing environments. In recent years, deep learning methods have achieved some major breakthroughs in the field of remote sensing image dehazing due to their powerful feature learning capabilities.

[0006] Deep learning methods have significantly improved dehazing performance due to their superior feature extraction capabilities and adaptability to complex scenarios. Deep learning methods can automatically extract features from hazy images through an end-to-end learning process, thus avoiding the limitations encountered by traditional methods. For example, convolutional neural networks (CNNs) acquire spatial features of images through numerous layers of convolutional operations, while generative adversarial networks (GANs) improve dehazing results through adversarial learning. These techniques have significant advantages in feature extraction and image reconstruction. Deep learning methods also excel in adapting to complex scenes. For instance, multi-scale convolutional neural networks fuse coarse-grained and fine-grained estimation networks, preserving more detailed information even in non-uniform fog distributions. The U-Net architecture, with its encoder-decoder structure, models global and local features of complex scenes, significantly improving the restoration of object edges and textures. These methods demonstrate stronger adaptability and robustness when dealing with remote sensing images containing various object types and complex terrains. It can be seen that deep learning methods have been widely used in research on dehazing algorithms for remote sensing images. They can cope with haze of different degrees and distributions and maintain the consistency of dehazing effect.

[0007] However, existing remote sensing image dehazing methods have certain drawbacks: ① Traditional dehazing methods struggle to accurately distinguish between low-frequency fog components caused by atmospheric scattering and inherent high-frequency detail features (such as edges and textures) in the frequency domain when processing remote sensing images. This leads to problems such as blurred details, residual noise, or artifacts during the dehazing process, especially noticeable in heavily foggy scenes. ② Existing technologies lack the ability to fuse multi-scale image features, relying on fixed rules or manually designed features. They cannot adaptively adjust the fusion strategy based on fog concentration and ground structure, resulting in limited representation capabilities of the fused features for complex scenes, making it difficult to achieve both dehazing effectiveness and detail preservation. ③ Deep learning methods heavily rely on large amounts of labeled data for training. Due to the high cost of acquiring and labeling remote sensing images... The data shortage is relatively high, the actual available dataset size is relatively small, and existing remote sensing image dehazing datasets are often only for certain specific tasks or scenarios, making it difficult to meet diverse needs; the shortage of data not only affects the model training effect, but also restricts the degree of generalization of the algorithm in different environments and scenarios; synthetic datasets alleviate the data shortage to some extent, but the fogged images they produce cannot completely imitate the complex fog distribution and lighting conditions in real scenes, making the model's generalization ability in practical use relatively poor; ④ Deep learning methods have relatively complex networks, a very large number of parameters, a long training time, and require high-performance hardware such as GPUs for support, which not only increases the research cost, but also limits the scope of application of the algorithm in practical applications. Summary of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of existing remote sensing image dehazing methods in terms of fog separation accuracy, detail preservation ability, adaptability and computational efficiency.

[0009] To address the aforementioned technical problems, this invention provides a remote sensing image dehazing method based on wavelet multi-scale decomposition, comprising: After performing three-level wavelet decomposition on the original foggy remote sensing image, low-frequency sub-band and three-level high-frequency sub-band are obtained; Nonlinear enhancement was performed on the LH subband, HL subband and HH subband in the three-layer high-frequency subband to obtain the three-layer enhanced high-frequency subband; A guide image is synthesized based on the three-layer enhanced high-frequency subbands, and then the low-frequency subbands are guided by the guide image to obtain the enhanced low-frequency subbands. Preliminary fusion features are obtained by fusing the enhanced low-frequency subband and the enhanced high-frequency subband with three layers using discrete wavelet inverse transform. The preliminary fusion features are then processed through dilated convolution, residual connection, and SE module to obtain the final fusion features. Based on fused features and three-layer enhanced high-frequency subbands, a three-layer decoding submodule is used to generate a dehazed remote sensing image. Each decoding submodule includes: fusing the input features of the current decoding submodule with the corresponding enhanced high-frequency subband after transposing convolution and bilateral filtering to obtain the detailed features of the current decoding submodule; applying the detailed features of the current decoding submodule to an adaptive sharpening kernel to obtain the output features of the current decoding submodule, which are then used as the input features of the next decoding submodule; the input features of the third decoding submodule are the fused features.

[0010] Preferably, the Daubechies-4 wavelet basis function is used to perform three-level wavelet decomposition on the original foggy remote sensing image.

[0011] Preferably, an S-shaped suppression enhancement function is used to nonlinearly enhance the LH subband in the three-layer high-frequency subband, as shown in the following formula: ; in, For the enhanced LH subband, For LH sub-band, The enhancement factor of the LH subband is set based on the local texture intensity. The location parameters of the LH sub-band are set according to the degree of haze obscuration.

[0012] Preferably, an exponential response enhancement function is used to nonlinearly enhance the HL subband in the three-layer high-frequency subband, as shown in the following formula: ; in, For the enhanced HL sub-band, For HL sub-band, The enhancement ratio of the HL sub-band. Enhanced response speed of HL subbands based on local texture intensity.

[0013] Preferably, a dual-gated nonlinear enhancement function is used to nonlinearly enhance the HH subband in the three-layer high-frequency subband, as shown in the following formula: ; in, For the enhanced HH subband, For HH subband, The enhancement ratio of the HH subband. and All of these are gating parameters.

[0014] Preferably, the low-frequency sub-band is subjected to multi-scale guided filtering based on the guiding image to obtain the enhanced low-frequency sub-band, including: Based on the guiding image, guide filtering is performed on the low-frequency subband using M windows with different radii, yielding the filtering results for the M windows, as shown in the formula: ; in, Let i be the filtering result of the m-th window, and i be the pixel index. For the m-th window, Let m be the number of pixels in the m-th window. The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. To guide the image, ; The adaptive weights for the filtering results of each window are calculated using the following formula: ; in, The adaptive weights for the m-th window. and The m-th window and the m-th window are respectively Scale-sensitive parameters for each window, The total number of windows, Let m be the local fog concentration at the m-th window. Low-frequency sub-band; The enhanced low-frequency sub-band is obtained based on the filtering results of each window and its adaptive weights, using the following formula: ; in, This is the enhanced low-frequency subband.

[0015] Preferably, the formulas for calculating the first and second linear coefficients of the m-th window are as follows: ; ; in, The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. For the m-th window, Let i be the number of pixels in the m-th window, and i be the pixel index. To guide the image, For low-frequency sub-band, Let be the mean value of the low-frequency sub-band within the m-th window. and Let be the mean and variance of the guiding image within the m-th window, respectively. This is the regularization parameter.

[0016] Preferably, the preliminary fused features are processed through dilated convolution, residual connections, and a SE module to obtain fused features, including: The initial fused features are input into three parallel dilated convolutional layers with different dilation rates to obtain three dilated convolutional output features; The initial fused features are added to the three dilated convolution output features respectively through residual connections to obtain three residual dilated convolution output features; The three residual dilated convolution output features are input into the SE module, the weight of each residual dilated convolution output feature is calculated, and the three residual dilated convolution output features are weighted and summed according to the weights to obtain the fused features.

[0017] Preferably, the detailed features of the current layer decoding submodule are processed by an adaptive sharpening kernel to obtain the output features of the current layer decoding submodule, which are then used as the input features of the next layer decoding submodule, including: The gradient magnitude of the detail features of the current layer decoding submodule is calculated using the following formula: ; in, For the detailed features of the j-th layer decoding submodule, and Let x and y be the gradient magnitudes of the detail features of the j-th layer decoding submodule, respectively. and They are respectively The convolution kernels of the operator in the x and y directions, Let be the gradient magnitude of the detail features of the j-th layer decoding submodule; ; The sharpening intensity of the current layer decoding submodule is calculated based on the gradient magnitude of its detailed features. The formula is as follows: ; in, Let be the sharpening intensity of the j-th layer decoding submodule. Scaling factor For the Sigmoid function, Strong and weak edge thresholds; The detail features are sharpened based on the sharpening intensity of the current layer's decoding submodule to obtain the output features of the current layer's decoding submodule, which are then used as the input features of the next layer's decoding submodule. The formula is as follows: ; in, For the input features of the (j-1)th layer decoding submodule, Apply Gaussian blur; The output characteristics of the last layer decoding submodule The image is a dehazed remote sensing image output by the decoding and reconstruction module.

[0018] The present invention also provides a remote sensing image dehazing device based on wavelet multi-scale decomposition, comprising: The multi-scale frequency decomposition module is used to perform three-level wavelet decomposition on the original foggy remote sensing image to obtain low-frequency sub-bands and three-level high-frequency sub-bands. The high-frequency detail dynamic enhancement module is used to perform nonlinear enhancement on the LH subband, HL subband and HH subband in the three-layer high-frequency subband, respectively, to obtain the three-layer enhanced high-frequency subband; The low-frequency sensing and guiding filter module is used to synthesize a guiding image based on the three-layer enhanced high-frequency sub-band, and then perform guiding filtering on the low-frequency sub-band based on the guiding image to obtain the enhanced low-frequency sub-band. The multi-scale context aggregation module is used to fuse the enhanced low-frequency subband and the three-layer enhanced high-frequency subband using discrete wavelet inverse transform to obtain preliminary fused features; the preliminary fused features are then processed through dilated convolution, residual connection and SE module to obtain the fused features; The decoding and reconstruction module is used to generate a dehazed remote sensing image based on fused features and three-layer enhanced high-frequency subbands, using three decoding sub-modules. Each decoding sub-module includes: fusing the input features of the current decoding sub-module with the corresponding enhanced high-frequency subband after transposing convolution and bilateral filtering to obtain the detailed features of the current decoding sub-module; applying the detailed features of the current decoding sub-module to an adaptive sharpening kernel to obtain the output features of the current decoding sub-module, which are then used as the input features of the next decoding sub-module; the input features of the third decoding sub-module are the fused features.

[0019] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a remote sensing image dehazing method based on wavelet multi-scale decomposition. First, multi-scale wavelet frequency decomposition divides the original foggy remote sensing image into low-frequency and high-frequency sub-bands, effectively separating fog interference and image details to preserve high-frequency details during subsequent dehazing. Then, a nonlinear dynamic enhancement function is applied to the high-frequency sub-band to strengthen details. The enhanced high-frequency sub-band is then used to apply multi-scale guided filtering to the low-frequency sub-band, suppressing fog components in the image. This achieves excellent dehazing results in light / heavy fog scenarios, improving image clarity and visual quality. Furthermore, residual dilated convolution and the SE module integrate feature information at different scales, enhancing the perception of global structure and local details to obtain fused features, avoiding edge loss during downsampling. Finally, the fused features are decoded and reconstructed to restore a high-resolution dehazed remote sensing image. This invention not only effectively improves the clarity, structural integrity, and application adaptability of remote sensing images after recovery under light to heavy haze conditions, but also reduces computational consumption and improves algorithm execution speed. It can operate normally even under limited resources, providing a reliable data foundation for tasks such as remote sensing interpretation, target recognition, and environmental monitoring. Attached Figure Description

[0020] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a remote sensing image dehazing method based on wavelet multi-scale decomposition according to the present invention; Figure 2 This is a schematic diagram of wavelet decomposition of the original foggy remote sensing image; Figure 3 This is a flowchart of the multi-scale context aggregation module; Figure 4 This is a visualization of the dehazing effect of the method of the present invention on a public dataset, wherein... Figure 4 Column (a) in the image shows the original remote sensing image with fog. Figure 4 Column (b) in the image contains clearly labeled remote sensing images. Figure 4 Column (c) in the image is a dehazed remote sensing image. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0022] Remote sensing images often suffer from quality degradation due to atmospheric scattering, cloud cover, and other factors, leading to the loss of key information and reduced analytical accuracy. Remote sensing image dehazing technology, as a core component of remote sensing image quality improvement and information reconstruction, has irreplaceable engineering value in fields such as dynamic surface monitoring, disaster emergency response, ecological environment assessment, precision agriculture management, urban planning, and resource exploration. By eliminating degrading interference and restoring image details, it provides a high-quality data foundation for subsequent interpretation and applications.

[0023] In the field of computer vision, atmospheric scattering models are often used to describe the formation of degraded images, and their mathematical expression is: ; in, For degraded images, It is a clear image. This represents the global atmospheric light value. For the transmittance of the medium; from the known... China Resumption A and t are the targets for defogging.

[0024] The first term of the equation This is called direct attenuation, which describes the scene's radiance and its attenuation in the medium; the second term of the equation... This is called atmospheric light, which originates from previously scattered light and causes a shift in scene color. When the atmosphere is uniform, the medium transmittance t can be expressed as: ; in, The total scattering coefficient, The scene depth represents the scene brightness, which means that scene brightness decreases exponentially with depth.

[0025] Remote sensing image dehazing aims to restore image quality, enhance features, and improve the accuracy of subsequent applications. It has irreplaceable core value in fields such as ecological environment monitoring, disaster emergency response, and military reconnaissance.

[0026] Reference Figure 1 As shown, this invention provides a remote sensing image dehazing method based on wavelet multi-scale decomposition, comprising: S1: After performing three-level wavelet decomposition on the original foggy remote sensing image, low-frequency sub-band and three-level high-frequency sub-band are obtained.

[0027] The goal of this invention is to create a remote sensing image dehazing algorithm based on wavelet feature enhancement. Leveraging the multi-scale decomposition capabilities of wavelet transform, it achieves an optimal balance between dehazing effectiveness and detail preservation. Wavelet transform is a time-frequency local analysis technique that divides an image into low-frequency and high-frequency sub-bands. The low-frequency sub-band represents the overall structure of the image, while the high-frequency sub-band corresponds to its finer details. By enhancing high-frequency features and suppressing low-frequency noise within the wavelet domain, it is possible to remove haze interference while preserving as much detail and color information as possible from the original image. This invention maintains the advantages of wavelet multi-scale decomposition and further enhances image clarity and resolution through feature enhancement.

[0028] This invention proposes a multi-scale frequency decomposition (MFD) module: The original hazy remote sensing image is input into the MFD module, which performs three-level wavelet decomposition on the image to obtain low-frequency sub-bands and three-level high-frequency sub-bands. This module fully utilizes the localization advantages of wavelet transform in the time and frequency domains, effectively dividing the remote sensing image into low-frequency and high-frequency sub-bands through multi-level decomposition. The low-frequency sub-bands carry the image's structural and brightness information, while the high-frequency sub-bands contain detailed features such as edges and textures.

[0029] Preferably, the present invention uses the Daubechies-4 (db4) wavelet basis function to perform discrete wavelet transform on the original foggy remote sensing image, which has tight support, high vanishing moment and good smoothness, and is suitable for preserving the spatial frequency variation of complex ground features in remote sensing images.

[0030] Daubechies wavelets are a family of orthogonal wavelets with compact support, orthogonality, and strong analytical power. db4 is a shorter member (with 4th-order filter coefficients) that offers the following significant advantages: (1) It has good time-frequency localization characteristics: The db4 wavelet has tight support characteristics in both the time and frequency domains, which can better preserve the local details of the image, and is especially suitable for processing local fog, clouds, edge structures and other detailed areas in remote sensing images.

[0031] (2) Strong edge preservation capability: The db4 wavelet has high vanishing moments, which is more conducive to data compression and noise elimination. It can effectively represent the smooth areas and abrupt boundaries of the image, capture edge structure and texture details at multiple scales, and facilitate structure restoration in subsequent dehazing enhancement.

[0032] (3) Balanced spatial and frequency resolution: Compared with shorter basis functions such as Haar wavelets, db4 can provide a smoother filtering response, avoiding excessive frequency leakage, and is suitable for processing rich mid-to-low frequency structural information in remote sensing images.

[0033] (4) Low computational complexity, suitable for engineering implementation: The db4 filter has a moderate length (supports a length of 8), which ensures both analysis accuracy and computational efficiency, making it suitable for the implementation requirements of real-time or semi-real-time remote sensing image processing systems.

[0034] In summary, choosing the db4 wavelet as the basic function for image wavelet decomposition can stably extract multi-scale features of images while preserving structural and detailed information, thus providing more accurate feature support for subsequent guided enhancement and image reconstruction.

[0035] Wavelet decomposition extracts spatial frequencies from an image using the Daubechies-4 wavelet basis function, dividing the image into low-frequency subbands (LL) and high-frequency subbands (LH, HL, HH). These four subbands are named according to a combination of "row frequency - column frequency," where: L (low): low-frequency components (corresponding to regions with low spatial frequencies); H (high): high-frequency components (corresponding to regions with high spatial frequencies); the first letter indicates the frequency component in the row direction (horizontal direction); the second letter indicates the frequency component in the column direction (vertical direction). Therefore, LL represents: row low frequency + column low frequency, indicating global low frequencies, describing the main structure and contours of the image; LH represents: row low frequency + column high frequency, indicating low frequencies in the horizontal direction and high frequencies in the vertical direction, describing the vertical edges / texture of the image; HL represents: row high frequency + column low frequency, indicating high frequencies in the horizontal direction and low frequencies in the vertical direction, describing the horizontal edges / texture of the image; HH represents: row high frequency + column high frequency, indicating global high frequencies, describing the details and noise of the image.

[0036] For the original foggy remote sensing image The image is decomposed into low-frequency subbands (LL) and high-frequency subbands (LH, HL, HH) using the MFD module. These components are used to extract image information at different scales. The low-frequency components contain global information of the image, while the high-frequency components preserve image details and edge features. This process is accomplished through continuous wavelet decomposition, using low-pass filters h[n-2k] and high-pass filters g[n-2k] in both row and column directions, respectively. After each filtering, a downsampling operation is performed, ultimately dividing the image into four subbands. The decomposed image can be effectively used for dehazing. The decomposition process uses the Daubechies-4 (db4) wavelet basis function as the core basis function for multi-scale frequency extraction.

[0037] Specifically, refer to Figure 2As shown, the original foggy remote sensing image is decomposed into low-frequency subbands and three high-frequency subbands using the Daubechies-4 wavelet basis function, including:

[0038] S11: Using the db4 wavelet basis function to process the original foggy remote sensing image Performing a Discrete Wavelet Transform (DWT) yields the result of the first-level wavelet decomposition, expressed as follows: ; in, This is the first low-frequency sub-band, representing the main structural information of the image; These are all first-layer high-frequency sub-bands, representing detailed information in the horizontal, vertical, and diagonal directions of the image, respectively. This is a discrete wavelet transform.

[0039] The first-level wavelet decomposition is detailed as follows: ; ; ; ; in, and These are low-pass and high-pass filters, respectively. The radius of the filter is denoted by "*". "*" indicates a convolution operation. This represents the original input signal, i.e., the original foggy remote sensing image. Represents the discrete location or index of data; and These represent row-by-row and column-by-column downsampling operations, respectively. In this embodiment, sampling is performed with a fixed step size of 2, retaining only pixel values ​​at even or odd positions to reduce image resolution while preserving specific frequency components.

[0040] S12: After completing the first-level wavelet decomposition, perform a second-level wavelet decomposition on the first-level low-frequency subband; then perform a third-level wavelet decomposition on the second-level low-frequency subband to obtain finer-grained low-frequency and high-frequency subbands. The second and third-level wavelet decompositions are represented as follows: ; in, and These are the low-frequency sub-bands of the j-th and (j-1)-th layers, respectively; All are high-frequency subbands of the j-th layer; .

[0041] The wavelet decomposition of the second and third layers is detailed below: ; ; ; ; in, This represents the low-frequency subband of the (j-1)th layer.

[0042] After performing three-level wavelet decomposition on the original foggy remote sensing image, a low-frequency subband and three-level high-frequency subbands are obtained, which can be represented by the MFD module as follows: ; in, This is the original remote sensing image with fog. This is a multi-scale frequency decomposition module. This is the third low-frequency sub-band. All are high-frequency subbands of the j-th layer.

[0043] Subsequent use Indicates the low-frequency sub-band. This indicates a high-frequency subband.

[0044] The advantages of the MFD module in dehazing remote sensing images are reflected in two aspects: First, the low-frequency subband can be used to model the haze components caused by atmospheric scattering and suppress them in subsequent guided filtering; second, the high-frequency subband, through a subband separation enhancement strategy, can achieve fine detail restoration with orientation awareness, significantly improving image edge sharpness and texture fidelity. Compared with traditional single-scale methods, the MFD module achieves frequency domain decoupling between image structure and haze interference, laying a robust foundation for subsequent feature enhancement and dehazing reconstruction, and exhibiting superior sharpness restoration capability and detail preservation performance under both light and heavy haze conditions.

[0045] S2: Nonlinear enhancement is performed on the LH subband, HL subband and HH subband in the three-layer high-frequency subband to obtain the three-layer enhanced high-frequency subband.

[0046] This invention proposes a High-frequency Detail Dynamic Enhancement (HDDE) module: three layers of high-frequency subbands are input into the HDDE module, and the LH subband, HL subband and HH subband in the high-frequency subband are nonlinearly enhanced respectively to obtain the three-layer enhanced high-frequency subbands.

[0047] Specifically, an S-shaped suppression enhancement function is used to nonlinearly enhance the horizontal LH subband in the three-layer high-frequency subband. This suppresses the response of small-amplitude coefficients, preventing flat regions from being amplified; and gradually enhances large-amplitude coefficients, which helps to preserve clear horizontal edges. The formula is: ; in, For the enhanced LH subband, For LH sub-band; The enhancement ratio of the LH subband is set according to the local texture intensity, and is used to control the overall enhancement ratio of the LH subband in the horizontal direction; The position parameters of the LH sub-band are set according to the degree of haze obscuration, and are used for horizontal control. The "saturation position" of the subband enhancement function.

[0048] An exponential response enhancement function is used to nonlinearly enhance the HL subband in a three-layer high-frequency subband. This method exhibits fast nonlinear response capabilities, significantly enhancing areas with strong vertical edges, such as those around buildings, while exhibiting relatively low noise response. The formula is: ; in, For the enhanced HL sub-band, For HL sub-band, The enhancement ratio of the HL sub-band is used to control the overall enhancement ratio of the HL sub-band in the vertical direction; The HL subband is set to enhance response speed based on local texture intensity. The larger this value is, the faster the function responds with small inputs, making it suitable for rapid edge enhancement.

[0049] A dual-gated nonlinear enhancement function is used to nonlinearly enhance the HH subband in the three-layer high-frequency subband. By introducing two gating functions with different sensitivities, the enhancement response to weak textures and large-scale diagonal edges is controlled respectively, adapting to complex texture regions and mixed directional signals. The formula is: ; in, For the enhanced HH subband, For HH subband, This is the enhancement ratio of the HH subband, used to control the overall enhancement ratio of the HH subband in the diagonal direction; and These are all gating parameters, which are adjusted based on the local fog concentration. A smaller value is used to control subtle texture enhancement, with a wider activation range. A relatively large value is used to control the edge activation of strong structures.

[0050] In general, this step It is an enhanced "magnification"; It is a regulator of "when to start strengthening"; The main adjustment is "Enhance Speed"; It refers to "gated response speed." It can combine multi-scale responses to achieve adaptive adjustment of image content, improving applicability and robustness.

[0051] This step ultimately yields the enhanced high-frequency subband, represented as follows: .

[0052] The HDDE module aims to enhance edge structures and texture details in remote sensing images that are easily obscured by fog by applying direction-aware nonlinear enhancement to three high-frequency sub-bands (LH, HL, and HH). Based on the frequency characteristics and ground feature distribution of each sub-band, the module designs differentiated adaptive enhancement functions to strengthen detail responses in the horizontal, vertical, and diagonal directions. By introducing sigmoid suppression, exponential response enhancement, and dual-gated nonlinear functions, the HDDE module achieves precise protection of weak textures and significant enhancement of strong edges, balancing detail preservation and noise suppression. Unlike traditional uniform enhancement strategies, this module adaptively adjusts the enhancement intensity based on prior information such as local texture intensity and fog obscuration level, significantly improving image clarity and structural integrity under varying fog concentrations. This module not only enhances weak but crucial edge features in remote sensing images but also provides richer and more reliable detail support for subsequent low-frequency guided filtering and fusion reconstruction, thereby improving the overall visual quality and information fidelity of dehazed images.

[0053] S3: Synthesize a guide image based on the three-layer enhanced high-frequency sub-bands, and then perform guide filtering on the low-frequency sub-bands based on the guide image to obtain the enhanced low-frequency sub-bands.

[0054] This invention proposes a Low-frequency-aware Guided Filter (LGF) module: A low-frequency sub-band and a three-layer enhanced high-frequency sub-band are input into the LGF module. A guiding image is synthesized based on the three-layer enhanced high-frequency sub-band. Then, the low-frequency sub-band is guided-filtered based on the guiding image to obtain the enhanced low-frequency sub-band. This module is specifically designed for the characteristic that fog components in remote sensing images are mainly concentrated in the low-frequency region. By introducing high-frequency details as a guiding image, it achieves effective suppression of fog in the low-frequency sub-band and explicit recovery of structural information.

[0055] First, the three-layer enhanced high-frequency subband The fused detail image serves as the original low-frequency subband. Guide image The formula is expressed as: ; in, To guide the image, This represents the enhanced high-frequency subband of the j-th layer.

[0056] The goal of guided filtering is to solve for the filtered output. Its essence is to establish and Local linear relationships, specifically including: For each pixel In the The central window Inside, assuming the output With guide map Satisfies a linear relationship: ; in, For the enhanced low-frequency subband, i is the pixel index. For window The first linear coefficient within, For window The second linear coefficient within, For windows.

[0057] Solve by minimizing the cost function. and : ; in, For low-frequency sub-band, For low-frequency subband in pixels The value; This is a regularization parameter used to prevent the image from becoming too smooth.

[0058] Solving the above cost function yields and : ; in, For window Number of pixels within, and Guide images In the window Mean and variance within, Low-frequency subband In the window The mean within.

[0059] Preferably, in this embodiment, multi-scale guided filtering is performed on the low-frequency sub-band based on the guiding image to obtain an enhanced low-frequency sub-band. This involves independently filtering using multiple windows of different radii, and then adaptively fusing the filtering results from each window. Specifically, this includes: S31: Based on the guiding image, guide filtering is performed on the low-frequency sub-band using M windows with different radii, yielding the filtering results for the M windows. The formula is as follows: ; in, Let i be the filtering result of the m-th window, and i be the pixel index. For the m-th window, Let m be the number of pixels in the m-th window. The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. To guide the image, .

[0060] S32: Calculate the adaptive weights of the filtering results for each window, using the following formula: ; in, The adaptive weights for the m-th window. and The m-th window and the m-th window are respectively Scale-sensitive parameters for each window, The total number of windows, Let m be the local fog concentration at the m-th window. Low-frequency sub-band; Scale-sensitive parameters ,in Let m be the window radius of the m-th window. It is a constant.

[0061] S33: The enhanced low-frequency sub-band is obtained based on the filtering results of each window and its adaptive weights, using the following formula: ; in, This is the enhanced low-frequency subband.

[0062] Specifically, the first linear coefficient of the m-th window Second linear coefficient The calculation formula is: ; ; in, The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. For the m-th window, Let i be the number of pixels in the m-th window, and i be the pixel index. To guide the image, For low-frequency sub-band, Let be the mean of the low-frequency subband within the m-th window. and Let be the mean and variance of the guiding image within the m-th window, respectively. This is the regularization parameter.

[0063] Compared to single-scale guided filtering, which uses a window with a fixed radius for filtering, the multi-scale guided filtering proposed in this embodiment has the following significant advantages: (1) Strong structural preservation: By filtering under multiple receptive fields, the overall contour and edge structure of the image can be extracted more fully, effectively alleviating the structural breakage problem caused by fog occlusion at small scale; (2) Good adaptability: Introducing an adaptive weight fusion function It adaptively adjusts the multi-scale response based on the local brightness difference at each pixel, effectively avoiding over-smoothing or over-enhancement caused by improper scale selection; (3) Strong noise suppression capability: This method can control the propagation of high-frequency noise while preserving the image structure perception capability, and avoid the introduction of texture artifacts when enhancing edges by traditional guided filtering; (4) Strong coupling with subsequent modules: the weighted low-frequency estimation results output It can better integrate low-frequency brightness information with high-frequency structure perception information, providing a clearer, more natural, and structurally accurate image foundation for subsequent operations, and participates in the overall reconstruction of dehazed images as a structure-guided signal, thereby improving the overall restoration quality.

[0064] The LGF module performs local linear guided filtering at multiple scales, adaptively adjusting filter weights to accommodate differences in fog concentration across different regions. This effectively removes blur and occlusion in low-frequency components while preserving the outlines and hierarchical structure of ground features. Compared to traditional single-scale filtering methods, the LGF module exhibits stronger structure preservation, adaptability, and noise suppression performance, avoiding detail loss and edge blurring caused by excessive smoothing. In the entire dehazing method, the LGF module is positioned after high-frequency detail enhancement and before feature fusion, undertaking the core task of bridging fog suppression and structure enhancement: on the one hand, it removes fog residue by accurately modeling low-frequency brightness information; on the other hand, it provides a clearer, higher-contrast structural foundation for subsequent feature fusion, significantly improving the clarity and naturalness of the final dehazed image. It is a key link in achieving coordinated optimization of frequency domain decoupling and spatial reconstruction.

[0065] In summary, by introducing the LGF module, this invention effectively solves the problems of structural blurring, edge blurring, and loss of detail caused by local fog or low-contrast areas in traditional low-frequency component processing. This module performs guided filtering operations at multiple scale windows and adaptively fuses the outputs from each scale based on local fog differences, thereby achieving enhanced perception and detail restoration of low-frequency structural information in the image.

[0066] S4: The low-frequency subband enhanced by inverse discrete wavelet transform and the high-frequency subband enhanced by three layers are fused to obtain the preliminary fusion features; the preliminary fusion features are then processed by dilated convolution, residual connection and SE module to obtain the fusion features.

[0067] This invention proposes a multi-scale contextual aggregation (MCA) module: Enhanced low-frequency subbands and three-layer enhanced high-frequency subbands are input into the MCA module, which then fuses them using inverse discrete wavelet transform to obtain preliminary fused features. These preliminary fused features are then processed through dilated convolution, residual connections, and a channel attention (SE) module to obtain the final fused features. This module aims to fuse low-frequency structural information and high-frequency detail information extracted during remote sensing image dehazing, enhancing the contextual relationships between features at different scales through residual dilated convolution and channel attention mechanisms.

[0068] Specifically, this step fuses the information extracted from the low-frequency subband and the high-frequency subband. Skip connections and context coding structures enable the feature fusion module to establish connections between shallow and deep layers, retain detailed feature information, and prevent the loss of edges or textures caused by downsampling. Furthermore, it fuses a multi-scale residual dilated convolution module, which, based on feature acquisition under different receptive fields, obtains contextual information that is crucial for dehazing remote sensing images.

[0069] Reference Figure 3 As shown, the specific steps of S4 include: S41: Preliminary fusion features are obtained by fusing the enhanced low-frequency subband and the enhanced high-frequency subband after three layers using discrete wavelet inverse transform.

[0070] Low-frequency subband and the three-layer enhanced high-frequency subband A preliminary fusion is performed using the Inverse Discrete Wavelet Transform (IDWT) with db4 as the inverse transform basis function to combine the information from both methods. The formula is as follows: ; in, As a preliminary fusion feature, This is the inverse discrete wavelet transform.

[0071] S42: The initial fused features are input into three parallel dilated convolutional layers with different dilation rates for multi-scale feature extraction, resulting in three dilated convolutional output features, as shown in the formula: ; in, For the first Each dilated convolution outputs features. For the Sigmoid function; For the expansion rate The dilated convolution operation inserts elements between the input feature map. To achieve expanded sampling, use zeros. Corresponding to different scales of holed convolutional layer branches; and These are the weights and biases of the dilated convolutional layer, respectively.

[0072] In the specific implementation, parameters and It can be shared between dilated convolutional layer branches at all scales, or it can be set independently for each scale of dilated convolutional layer branches, thus balancing lightweightness and expressive power respectively.

[0073] S43: The preliminary fused features are added to the three dilated convolution output features respectively through residual connections to obtain three residual dilated convolution output features, expressed by the formula: ; in, For the first Each residual dilated convolution outputs a feature.

[0074] The core idea of ​​residual connections is to add the input features (or their feature transformations) to the features after nonlinear transformation, which helps to smoothly propagate gradients and preserve important features. Residual dilated convolution captures multi-scale fog distribution features, channel attention dynamically allocates sub-band weights, and finally achieves synergistic optimization of dehazing and detail through weighted summation.

[0075] S44: Input the three residual dilated convolution output features into the SE (Squeeze-and-Excitation) module, calculate the weight of each residual dilated convolution output feature, and sum the three residual dilated convolution output features according to the weights to obtain the fused features.

[0076] The SE module outputs features through a "compression-excitation" mechanism by performing dilated convolutions on the three input residuals. Adaptive weighting is performed to enhance the expressive power of important subbands in the final fusion. The specific steps are as follows: S44-1: Channel Compression (Squeeze): Output features for each residual dilated convolution. Perform global average pooling to calculate the channel vector, using the following formula: ; in, The first channel vector One element, and These represent the height and width of the residual dilated convolution output feature, respectively. according to Obtain the channel vector .

[0077] S44-2: Channel Excitation: This involves applying the channel vector... The input is fed into a two-layer fully connected network to model the nonlinear relationship between channels, calculate the weights of the output features of each residual dilated convolution, and generate normalized weights. ,in For the first The weights of the residual dilated convolution output features. The specific formula is: ; in, To normalize the weights, The weight matrix is ​​for a two-layer fully connected system. This is the compression ratio (default is set to 2 or 4). The Sigmoid function is used to normalize the output. interval; This is the ReLU activation function.

[0078] S44-3: Channel Reweighting: Based on weights Output features of residual dilated convolution Weighted summation yields the fusion characteristics, enabling channel recalibration and weighted summation. The formula is as follows: ; in, This is a feature of fusion.

[0079] By introducing the channel attention mechanism of the SE module, the responsiveness of key features can be effectively improved, the direction awareness and structure preservation capabilities during the dehazing process can be enhanced, and fused features with enhanced high-frequency detail information and low-frequency structural information fused at multiple scales can be obtained. This serves as the input for the subsequent decoding and reconstruction stage.

[0080] The MCA module first utilizes dilated convolution to capture multi-scale spatial semantic information of images under different receptive fields, making it particularly suitable for complex scenes in remote sensing images with varying ground feature scales and uneven fog distribution. Subsequently, it introduces an SE channel attention mechanism to adaptively adjust the importance weights of each sub-band, highlighting key structures and edge details, effectively mitigating texture loss caused by downsampling. In the entire dehazing system, the MCA module, as the core fusion unit connecting the frequency domain enhancement (HDDE, LGF) and decoding / reconstruction modules, not only integrates multi-source information to form a unified feature representation but also improves the compatibility between global consistency and local finesse, ensuring that the final restored image possesses high structural fidelity, clarity, and naturalness. It is a crucial hub for achieving multi-scale information collaborative optimization and full-chain enhancement and reconstruction.

[0081] S5: Based on the fused features and the enhanced high-frequency subbands of the three layers, a dehazed remote sensing image is generated using a three-layer decoding submodule. Each decoding submodule includes: fusing the input features of the current layer decoding submodule with the enhanced high-frequency subbands of the corresponding layer after transposed convolution and bilateral filtering to obtain the detailed features of the current layer decoding submodule; applying the detailed features of the current layer decoding submodule to an adaptive sharpening kernel to obtain the output features of the current layer decoding submodule, which are then used as the input features of the next layer decoding submodule; the input features of the third layer decoding submodule are the fused features.

[0082] The decoding and reconstruction module for detail restoration is the final output of the method of this invention. It is mainly responsible for restoring the fused features after multi-scale fusion into a high-resolution, clear remote sensing image, while enhancing edges and textures through detail injection and adaptive sharpening.

[0083] Specifically, the fused features are input to the third-layer decoding submodule and fused with the enhanced high-frequency subband of the third layer to obtain the third-layer decoded features. These third-layer decoded features then sequentially pass through the second-layer and first-layer decoding submodules, respectively, and are fused with the enhanced high-frequency subbands of the second and first layers to obtain the dehazed remote sensing image. The specific steps of each decoding submodule include: S51: The input features of the current layer's decoding submodule are subjected to transposed convolution to obtain upsampled features, expressed by the formula: ; in, For the input features of the j-th layer decoding submodule, This indicates the transpose convolution operation. For the upsampling features of the j-th layer decoding submodule; This is a hierarchical index for the wavelet decomposition and decoding submodules.

[0084] Specifically, the input features of the third-layer decoding submodule are fused features. The upsampled features of the third-layer decoding submodule are obtained. Similarly, the upsampling features of the second-layer decoding submodule are: The upsampling features of the first-layer decoding submodule are , and These represent the height and width of the original foggy remote sensing image, respectively.

[0085] By employing transposed convolution with a stride of 2, a kernel size of 4×4, and 'same' padding, upsampling is performed to ensure that image detail is preserved to the greatest extent while restoring image resolution, thus generating a dehazed high-resolution remote sensing image.

[0086] S52: To suppress artifacts introduced by transposed convolution, the upsampled features of the current layer's decoding submodule are subjected to bilateral filtering, preserving edges while smoothing noise, to obtain the initial detail features of the current layer's decoding submodule, expressed by the formula: ; in, These are the initial detail features of the j-th layer decoding submodule; The normalization coefficient is... For The neighborhood centered on, Center pixel coordinates The pixel coordinates within the domain. For the kernel of the bilateral filter, For spatial distance, This represents the difference in pixel values.

[0087] Transposed convolution and bilateral filtering progressively restore the resolution of the input features of the current layer's decoding submodule to the input size of the original remote sensing image.

[0088] S53: The initial detail features of the current layer's decoding submodule are fused with the corresponding enhanced high-frequency subband of the current layer. The high-frequency subband is used to supplement the edge and texture details lost due to downsampling, thus obtaining the detail features of the current layer's decoding submodule. The formula is as follows: ; in, For the detailed features of the j-th layer decoding submodule, The output features of the bilateral filter in the j-th layer decoding submodule are... These are learnable weights for the decoding submodule, used to balance the contributions of fused features and high-frequency details; This is the inverse discrete wavelet transform. This represents the enhanced high-frequency subband of the j-th layer.

[0089] S54: To further enhance image edges, an adaptive sharpening kernel is designed based on local gradient magnitude. The detailed features of the current layer decoding submodule are then processed by the adaptive sharpening kernel to obtain the output features of the current layer decoding submodule, which are then used as the input features of the next layer decoding submodule, including: The gradient magnitude of the detail features of the current layer decoding submodule is calculated using the following formula: ; in, For the detailed features of the j-th layer decoding submodule, and Let x and y be the gradient magnitudes of the detail features of the j-th layer decoding submodule, respectively. and They are respectively The convolution kernels of the operator in the x and y directions, Let be the gradient magnitude of the detail features of the j-th layer decoding submodule; The sharpening intensity of the current layer decoding submodule is calculated based on the gradient magnitude of its detailed features. The formula is as follows: ; in, Let be the sharpening intensity of the j-th layer decoding submodule. This is a scaling factor used to control the range of sharpening intensity; The Sigmoid function is used to map gradients to the [0,1] interval; This is the threshold for strong and weak edges, used to distinguish between strong and weak edge regions; The detail features are sharpened based on the sharpening intensity of the current layer's decoding submodule to obtain the output features of the current layer's decoding submodule, which are then used as the input features of the next layer's decoding submodule. The formula is as follows: ; in, For the input features of the (j-1)th layer decoding submodule, Gaussian blur (kernel size 3×3, standard deviation 1); This is the Laplace operator approximation.

[0090] The output characteristics of the last layer decoding submodule The formula for the dehazed remote sensing image output by the decoding and reconstruction module is: ; in, This is a dehazed remote sensing image.

[0091] The decoding and reconstruction module achieves progressive upsampling through multi-level transposed convolutions and introduces bilateral filtering to eliminate artifacts that may be introduced during upsampling. Simultaneously, it injects enhanced high-frequency detail subbands through skip connections to further recover edges and textures lost due to downsampling or haze occlusion. To further enhance image sharpness and depth, the module also incorporates an adaptive sharpening mechanism based on local gradient information to achieve precise enhancement of strong edge regions. The advantages of this module are twofold: firstly, it ensures the clarity and structural integrity of the image reconstruction process, avoiding the blurring and distortion problems common in traditional decoding; secondly, through close cooperation with the frequency domain enhancement module (HDDE, LGF) and the multi-scale fusion module (MCA), it achieves full-chain collaborative optimization from frequency domain decomposition to spatial restoration. In the entire system, the decoding and reconstruction module not only undertakes the final task of image generation but is also a key node in effectively transforming the results of previous feature mining and enhancement into visible image quality improvement, directly determining the practicality and visual performance of the final remote sensing image.

[0092] To verify the effectiveness of the method of this invention, this embodiment also selects the publicly available remote sensing image dataset StateHaze1k as the experimental data source. This dataset consists of 1200 pairs of clear synthetic aperture radar (SAR) images, foggy visible light remote sensing images, and corresponding ground real images. Based on the different levels of fog concentration, i.e., the degree of degradation of the remote sensing images, it is divided into three sets: light fog (StateHaze1k-TN), moderate fog (StateHaze1k-M), and dense fog (StateHaze1k-TK), each containing 400 image pairs. In the synthetic images covered by light fog, the fog mask is only the fog extracted from the original real cloudy image. For the moderate fog images, the samples are a superposition of fog and moderately concentrated fog. For the dense fog images, the transmission map is selected from the dense fog.

[0093] The original foggy remote sensing image was set to a size of 512×512. The colors representing different land cover categories in the RGB labels were mapped to their corresponding grayscale values. The image and labels were cropped to meet the model's input size requirements. The cropped image was then uniformly sized to 512×512 pixels. The cropped image was then input into the deep network for training.

[0094] Three sets of 400 sample images were randomly selected, with 350 images randomly chosen from each set for training the deep network model and 30 images from each set for validation. After training, all remaining sample images were input into the deep network model for testing, and the dehazing results of the test samples were obtained. Peak signal-to-noise ratio and structural similarity were used to evaluate the classification results.

[0095] The dehazing results obtained using the remote sensing image dehazing method proposed in this invention, traditional methods, and deep learning methods are shown in Table 1.

[0096] Table 1. Comparison of Experimental Results

[0097] As can be seen from the table, in a light haze scenario, the method of this invention achieves peak signal-to-noise ratio (PSNR) of 32.35dB and structural similarity (SSIM) of 0.925, which are 3.5dB and 0.065 higher than the traditional method (28.85dB, 0.86), and 1.35dB and 0.015 higher than the deep learning method (31dB, 0.91), respectively, indicating a significant advantage in detail recovery and structure preservation.

[0098] In heavy haze scenarios, the PSNR and SSIM of the method of this invention are 30.89 dB and 0.968, respectively, which are significantly improved compared to the traditional method (26.95 dB, 0.825), by 3.94 dB and 0.143, respectively; compared to the deep learning method (28.9 dB, 0.88), they are also improved by 1.99 dB and 0.088, respectively. This fully demonstrates that the present invention can maintain good image reconstruction capability and structural fidelity under high haze conditions.

[0099] Figure 4 This is a visualization of the dehazing effect of the method of the present invention on a public dataset, wherein... Figure 4 Column (a) in the image shows the original remote sensing image with fog. Figure 4 Column (b) in the image contains clearly labeled remote sensing images. Figure 4 Column (c) in the image is a dehazed remote sensing image.

[0100] It can be seen that the wavelet feature-enhanced remote sensing image dehazing algorithm proposed in this invention has achieved good results in both light and heavy haze scenarios. Compared with traditional methods and deep learning methods, this invention has achieved significant improvements in PSNR and SSIM metrics, indicating that the algorithm has obvious advantages in improving image clarity and preserving image details.

[0101] Therefore, the method of this invention achieves quantitative performance superior to existing methods under different haze intensities, especially exhibiting stronger robustness and recovery ability in heavy fog scenarios, verifying the significant effectiveness of the designed wavelet multi-scale enhancement and detail guidance mechanism in remote sensing image dehazing tasks.

[0102] In summary, the remote sensing image dehazing method based on wavelet multi-scale decomposition described in this invention first divides the original foggy remote sensing image into low-frequency and high-frequency sub-bands through multi-scale wavelet frequency decomposition, effectively separating fog interference and image details to fully preserve high-frequency details in the subsequent dehazing process. Then, a nonlinear dynamic enhancement function is applied to the high-frequency sub-band to strengthen details, and the enhanced high-frequency sub-band is used to perform multi-scale guided filtering on the low-frequency sub-band to suppress fog components in the image, achieving excellent dehazing effects in light / heavy fog scenarios and improving image clarity and visual quality. Furthermore, residual dilated convolution and the SE module are used to integrate feature information at different scales, enhancing the perception of global structure and local details to obtain fused features, avoiding edge loss during downsampling. Finally, the fused features are decoded and reconstructed to recover a high-resolution dehazed remote sensing image. This invention not only effectively improves the clarity, structural integrity, and application adaptability of remote sensing images after recovery under light to heavy haze conditions, but also reduces computational consumption and improves algorithm execution speed. It can operate normally even under limited resources, providing a reliable data foundation for tasks such as remote sensing interpretation, target recognition, and environmental monitoring.

[0103] Based on the aforementioned remote sensing image dehazing method based on wavelet multi-scale decomposition, this invention also provides a remote sensing image dehazing device based on wavelet multi-scale decomposition, comprising: The multi-scale frequency decomposition module is used to perform three-level wavelet decomposition on the original foggy remote sensing image to obtain low-frequency sub-bands and three-level high-frequency sub-bands. The high-frequency detail dynamic enhancement module is used to perform nonlinear enhancement on the LH subband, HL subband and HH subband in the three-layer high-frequency subband, respectively, to obtain the three-layer enhanced high-frequency subband; The low-frequency sensing and guiding filter module is used to synthesize a guiding image based on the three-layer enhanced high-frequency sub-band, and then perform guiding filtering on the low-frequency sub-band based on the guiding image to obtain the enhanced low-frequency sub-band. The multi-scale context aggregation module is used to fuse the enhanced low-frequency subband and the three-layer enhanced high-frequency subband using discrete wavelet inverse transform to obtain preliminary fused features; the preliminary fused features are then processed through dilated convolution, residual connection and SE module to obtain the fused features; The decoding and reconstruction module is used to generate dehazed remote sensing images based on fused features and three layers of enhanced high-frequency subbands using three decoding sub-modules. Each decoding sub-module includes: fusing the input features of the current decoding sub-module with the corresponding enhanced high-frequency subband after transposing convolution and bilateral filtering to obtain the detailed features of the current decoding sub-module; applying the detailed features of the current decoding sub-module to an adaptive sharpening kernel to obtain the output features of the current decoding sub-module, which are then used as the input features of the next decoding sub-module; the input features of the third decoding sub-module are the fused features.

[0104] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for dehazing remote sensing images based on wavelet multi-scale decomposition, characterized in that, include: After performing three-level wavelet decomposition on the original foggy remote sensing image, low-frequency sub-band and three-level high-frequency sub-band are obtained; Nonlinear enhancement was performed on the LH subband, HL subband and HH subband in the three-layer high-frequency subband to obtain the three-layer enhanced high-frequency subband; A guide image is synthesized based on the three-layer enhanced high-frequency subbands, and then the low-frequency subbands are guided by the guide image to obtain the enhanced low-frequency subbands. Preliminary fusion features are obtained by fusing the enhanced low-frequency subband and the enhanced high-frequency subband with three layers using discrete wavelet inverse transform. The initial fused features are processed through dilated convolution, residual connections, and SE modules to obtain the fused features; Based on the fusion features and the high-frequency subband enhanced by three layers, a dehazed remote sensing image is generated using a three-layer decoding submodule. Each decoding submodule includes: transposing the input features of the current decoding submodule and performing bilateral filtering, then fusing them with the enhanced high-frequency subbands of the current layer to obtain the detailed features of the current decoding submodule; applying the detailed features of the current decoding submodule to an adaptive sharpening kernel to obtain the output features of the current decoding submodule, which are then used as the input features of the next decoding submodule; the input features of the third decoding submodule are the fused features.

2. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The original foggy remote sensing image was decomposed into three levels using the Daubechies-4 wavelet basis function.

3. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The LH subband in the three-layer high-frequency subband is nonlinearly enhanced using the S-shaped suppression enhancement function, as shown in the formula: ； in, For the enhanced LH subband, For LH sub-band, The enhancement factor of the LH subband is set based on the local texture intensity. The location parameters of the LH sub-band are set according to the degree of haze obscuration.

4. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The exponential response enhancement function is used to nonlinearly enhance the HL subband in the three-layer high-frequency subband, as shown in the following formula: ； in, For the enhanced HL sub-band, For HL sub-band, The enhancement ratio of the HL sub-band. Enhanced response speed of HL subbands based on local texture intensity.

5. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The HH subband in the three-layer high-frequency subband is nonlinearly enhanced using a dual-gated nonlinear enhancement function, as shown in the following formula: ； in, For the enhanced HH subband, For HH subband, The enhancement ratio of the HH subband, and All of these are gating parameters.

6. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, Multi-scale guided filtering is performed on the low-frequency subband based on the guiding image to obtain the enhanced low-frequency subband, including: Based on the guiding image, guide filtering is performed on the low-frequency subband using M windows with different radii, yielding the filtering results for the M windows, as shown in the formula: ； in, Let i be the filtering result of the m-th window, and i be the pixel index. For the m-th window, Let m be the number of pixels in the m-th window. The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. To guide the image, ; The adaptive weights for the filtering results of each window are calculated using the following formula: ； in, The adaptive weights for the m-th window. and The m-th window and the m-th window are respectively Scale-sensitive parameters for each window, The total number of windows, Let m be the local fog concentration at the m-th window. For low-frequency sub-band; The enhanced low-frequency sub-band is obtained based on the filtering results of each window and its adaptive weights, using the following formula: ； in, This is the enhanced low-frequency subband.

7. A remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 6, characterized in that, The formulas for calculating the first and second linear coefficients of the m-th window are: ； ； in, The first linear coefficient of the m-th window. The second linear coefficient of the m-th window. For the m-th window, Let i be the number of pixels in the m-th window, and i be the pixel index. To guide the image, For low-frequency sub-band, Let be the mean of the low-frequency subband within the m-th window. and Let be the mean and variance of the guiding image within the m-th window, respectively. This is the regularization parameter.

8. The remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The initial fused features are processed through dilated convolution, residual connections, and SE modules to obtain fused features, including: The initial fused features are input into three parallel dilated convolutional layers with different dilation rates to obtain three dilated convolutional output features; The initial fused features are added to the three dilated convolution output features respectively through residual connections to obtain three residual dilated convolution output features; The three residual dilated convolution output features are input into the SE module, the weight of each residual dilated convolution output feature is calculated, and the three residual dilated convolution output features are weighted and summed according to the weights to obtain the fused features.

9. A remote sensing image dehazing method based on wavelet multi-scale decomposition according to claim 1, characterized in that, The detailed features of the current layer decoding submodule are processed by an adaptive sharpening kernel to obtain the output features of the current layer decoding submodule, which are then used as the input features of the next layer decoding submodule, including: The gradient magnitude of the detail features of the current layer decoding submodule is calculated using the following formula: ； in, For the detailed features of the j-th layer decoding submodule, and Let x and y be the gradient magnitudes of the detail features of the j-th layer decoding submodule, respectively. and They are respectively The convolution kernels of the operator in the x and y directions, Let be the gradient magnitude of the detail features of the j-th layer decoding submodule; ; The sharpening intensity of the current layer decoding submodule is calculated based on the gradient magnitude of its detailed features. The formula is as follows: ； in, Let be the sharpening intensity of the j-th layer decoding submodule. Scaling factor For the Sigmoid function, Strong and weak edge thresholds; The detail features are sharpened based on the sharpening intensity of the current layer's decoding submodule to obtain the output features of the current layer's decoding submodule, which are then used as the input features of the next layer's decoding submodule. The formula is as follows: ； in, For the input features of the (j-1)th layer decoding submodule, Apply Gaussian blur; The output characteristics of the last layer decoding submodule The image is a dehazed remote sensing image output by the decoding and reconstruction module.

10. A remote sensing image dehazing device based on wavelet multi-scale decomposition, characterized in that, include: The multi-scale frequency decomposition module is used to perform three-level wavelet decomposition on the original foggy remote sensing image to obtain low-frequency sub-bands and three-level high-frequency sub-bands. The high-frequency detail dynamic enhancement module is used to perform nonlinear enhancement on the LH subband, HL subband and HH subband in the three-layer high-frequency subband, respectively, to obtain the three-layer enhanced high-frequency subband; The low-frequency sensing and guiding filter module is used to synthesize a guiding image based on the three-layer enhanced high-frequency sub-band, and then perform guiding filtering on the low-frequency sub-band based on the guiding image to obtain the enhanced low-frequency sub-band. The multi-scale context aggregation module is used to fuse the enhanced low-frequency subband and the three-layer enhanced high-frequency subband using discrete wavelet inverse transform to obtain preliminary fused features; the preliminary fused features are then processed through dilated convolution, residual connection and SE module to obtain the fused features; The decoding and reconstruction module is used to generate a dehazed remote sensing image by using the high-frequency subband based on fusion features and three-layer enhancement after passing through the three-layer decoding submodule. Each decoding submodule includes: transposing the input features of the current decoding submodule and performing bilateral filtering, then fusing them with the enhanced high-frequency subbands of the current layer to obtain the detailed features of the current decoding submodule; applying the detailed features of the current decoding submodule to an adaptive sharpening kernel to obtain the output features of the current decoding submodule, which are then used as the input features of the next decoding submodule; the input features of the third decoding submodule are the fused features.

Images