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

Abstract

The present application 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, which first separates the original foggy remote sensing image into a low-frequency subband and a high-frequency subband through multi-scale wavelet frequency decomposition, separates the fog and haze interference and image details; then uses a nonlinear dynamic enhancement function to strengthen the details of the high-frequency subband, uses the enhanced high-frequency subband to perform multi-scale guided filtering on the low-frequency subband, and suppresses the fog component in the image; further integrates the feature information under different scales through residual hole convolution and SE module to strengthen the perception ability of global structure and local details, and obtains fused features; finally, the fused features are decoded and reconstructed to obtain a high-resolution defogging remote sensing image. The present application not only effectively improves the clarity, structural integrity and application adaptability of the recovered remote sensing image under light to heavy fog and haze conditions, but also reduces the operation consumption and improves the algorithm execution speed.

Description

TECHNICAL FIELD

[0001] The present application 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. BACKGROUND

[0002] Due to the influence of atmospheric interference such as haze, fog and thin clouds on remote sensing images, ground observation and subsequent target understanding and analysis tasks are greatly hindered. Remote sensing image defogging plays a crucial role in improving the data usability of applications such as environmental monitoring, disaster response and military reconnaissance.

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

[0004] Traditional defogging methods mostly rely on image enhancement and physical models, including Histogram Equalization (HE) and Dark Channel Prior (DCP). Histogram equalization is a traditional image enhancement method, whose core concept is to expand the distribution interval of gray values to enhance the overall contrast of the image. This technique has a significant effect on images with concentrated gray values or low contrast, and is very suitable for handling insufficient contrast between the background and foreground, but still has certain limitations in remote sensing image defogging. The dark channel prior method mainly uses the statistical properties of the dark channel in the haze-free image to estimate the transmittance and atmospheric light value of the haze image. By analyzing the intensity distribution of the dark channel in the haze image, the contribution of the atmospheric light is estimated, and then the transmittance is calculated to restore the haze-free image. When processing images with sky, the dark channel prior is prone to color distortion problems. Because the sky area often has a high brightness value, the dark channel prior will mistakenly regard the sky area as atmospheric light when estimating the transmittance, resulting in color shift. In areas with sudden changes in depth of field, such as where mountains meet the sky, the dark channel prior may also form halos, as this method does not take into account the differences in local features when estimating transmittance, resulting in loss of detail information in the boundary area.

[0005] Traditional remote sensing image defogging methods mostly rely on physical models or image enhancement methods, which can improve image quality to some extent, but have obvious shortcomings in detail preservation, color correction and adaptability to complex and variable environments. In recent years, deep learning methods have made some major breakthroughs in remote sensing image defogging due to their strong feature learning ability.

[0006] Deep learning methods have significantly improved the dehazing performance by virtue of their outstanding feature extraction ability and adaptability to complex scenes. Deep learning methods can automatically extract the characteristics of haze images through an end-to-end learning process, thereby avoiding the limitations encountered by traditional methods. For example, convolutional neural networks obtain spatial features of images through multiple layers of convolution operations, and generative adversarial networks improve dehazing results by means of adversarial learning. These techniques have obvious advantages in feature extraction and image restoration. In terms of adaptability to complex scenes, deep learning methods also have their own unique advantages. For example, multi-scale convolutional neural networks combine coarse-grained and fine-grained estimation networks, which can preserve more detailed information in the case of non-uniform haze distribution. The U-Net architecture uses an encoder-decoder structure to model global and local features of complex scenes, thereby significantly improving the restoration of edges and textures of ground objects. These methods exhibit stronger adaptability and robustness when dealing with remote sensing images containing multiple ground object types and complex terrain. It can be seen that deep learning methods have been widely applied in the research of remote sensing image dehazing algorithms, and they can handle different levels and distributions of haze while maintaining the consistency of dehazing results.

[0007] However, the existing remote sensing image dehazing methods have certain drawbacks: ① Traditional dehazing methods have difficulty in accurately distinguishing between low-frequency haze components caused by atmospheric scattering and inherent high-frequency detail features (such as edges and textures) of images in the frequency domain when processing remote sensing images. This leads to problems such as blurring of details, residual noise, or artifacts during the dehazing process, especially in heavy haze scenarios. ② Existing techniques lack the ability to fuse multi-scale features of images, relying on fixed rules or manually designed features. They cannot adaptively adjust the fusion strategy based on haze density and ground object structure, resulting in limited representation ability of fused features for complex scenes and difficulty in balancing dehazing results and detail fidelity. ③ Deep learning methods heavily rely on large amounts of labeled data for training. Due to the high cost of obtaining and labeling remote sensing images, the actual available dataset size is relatively small. Moreover, existing remote sensing image dehazing datasets are often only for specific tasks or scenarios, making it difficult to meet diverse needs. The shortage of data not only affects the effectiveness of model training but also restricts the degree of algorithm generalization in different environments and scenarios. Synthetic datasets alleviate the data shortage to some extent, but the generated haze images cannot completely simulate complex haze distribution and lighting conditions in real scenes, resulting in poor generalization ability of models in actual use. ④ Deep learning methods have complex networks with a large number of parameters, requiring a long training time and high-performance GPUs and other hardware to support. This not only increases the research cost but also limits the application range of the algorithm in practical applications. SUMMARY

[0008] To this end, the technical problem to be solved by the present application is to overcome the problems of the existing remote sensing image defogging methods in terms of fog separation precision, detail preservation ability, adaptability and computational efficiency.

[0009] To solve the above technical problems, the present application provides a remote sensing image defogging method based on wavelet multi-scale decomposition, comprising:

[0010] After three-layer wavelet decomposition of the original foggy remote sensing image, a low-frequency subband and three high-frequency subbands are obtained.

[0011] The LH subband, HL subband and HH subband in the three high-frequency subbands are respectively subjected to nonlinear enhancement to obtain three enhanced high-frequency subbands.

[0012] A guided image is synthesized according to the three enhanced high-frequency subbands, and the low-frequency subband is guided filtered according to the guided image to obtain an enhanced low-frequency subband.

[0013] The enhanced low-frequency subband and the three enhanced high-frequency subbands are fused by using a discrete wavelet inverse transform to obtain a preliminary fusion feature; the preliminary fusion feature is subjected to a hole convolution, residual connection and SE module to obtain a fusion feature.

[0014] Based on the fusion feature and the three enhanced high-frequency subbands, a three-layer decoding sub-module is used to generate a defogging remote sensing image; each layer decoding sub-module comprises: after the input feature of the current layer decoding sub-module is subjected to transposed convolution and bilateral filtering, it is fused with the enhanced high-frequency subband corresponding to the current layer to obtain the detail feature of the current layer decoding sub-module; the detail feature of the current layer decoding sub-module is subjected to an adaptive sharpening kernel to obtain the output feature of the current layer decoding sub-module, which is used as the input feature of the next layer decoding sub-module; the input feature of the third layer decoding sub-module is the fusion feature.

[0015] Preferably, the Daubechies-4 wavelet basis function is used for three-layer wavelet decomposition of the original foggy remote sensing image.

[0016] Preferably, the LH subband in the three high-frequency subbands is subjected to nonlinear enhancement by using an S-shaped inhibition enhancement function, and the formula is:

[0017] ;

[0018] wherein, is the enhanced LH subband, is the LH subband, is an enhancement ratio of the LH subband set according to the local texture intensity, is a position parameter of the LH subband set according to the haze obscuring degree.

[0019] Preferably, the HL subband in the three-layer high-frequency subband is nonlinearly enhanced by using an exponential response enhancement function, and the formula is:

[0020] ;

[0021] wherein, is the enhanced HL subband, is the HL subband, is the enhancement ratio of the HL subband, is the enhancement response speed of the HL subband set according to the local texture intensity.

[0022] Preferably, the HH subband in the three-layer high-frequency subband is nonlinearly enhanced by using a double-gated nonlinear enhancement function, and the formula is:

[0023] ;

[0024] wherein, is the enhanced HH subband, is the HH subband, is the enhancement ratio of the HH subband, and are gate parameters.

[0025] Preferably, the low-frequency subband is multiscale guided filtered according to the guide image to obtain an enhanced low-frequency subband, including:

[0026] According to the guide image, the low-frequency subband is guided filtered by using M windows with different window radii to obtain M window filtering results, and the formula is:

[0027] ;

[0028] wherein, is the filtering result of the mth window, i is a pixel index, is the mth window, is the number of pixels in the mth window, is the first linear coefficient of the mth window, is the second linear coefficient of the mth window, is the guide image, ;

[0029] The adaptive weight of the filtering result of each window is calculated, and the formula is:

[0030] ;

[0031] wherein, is the adaptive weight of the mth window, and are the mth window and the nth window, 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;

[0032] The enhanced low-frequency sub-band is obtained based on the filtering results of each window and its adaptive weights, using the following formula:

[0033] ;

[0034] in, This is the enhanced low-frequency subband.

[0035] Preferably, the formulas for calculating the first and second linear coefficients of the m-th window are as follows:

[0036] ;

[0037] ;

[0038] 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.

[0039] Preferably, the preliminary fused features are processed through dilated convolution, residual connections, and a SE module to obtain fused features, including:

[0040] The initial fused features are input into three parallel dilated convolutional layers with different dilation rates to obtain three dilated convolutional output features;

[0041] 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;

[0042] 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.

[0043] 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:

[0044] The gradient magnitude of the detail features of the current layer decoding submodule is calculated using the following formula:

[0045] ;

[0046] 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; ;

[0047] 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:

[0048] ;

[0049] in, Let be the sharpening intensity of the j-th layer decoding submodule. Scaling factor For the Sigmoid function, Strong and weak edge thresholds;

[0050] 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:

[0051] ;

[0052] in, For the input features of the (j-1)th layer decoding submodule, Apply Gaussian blur;

[0053] The output characteristics of the last layer decoding submodule The image is a dehazed remote sensing image output by the decoding and reconstruction module.

[0054] The present invention also provides a remote sensing image dehazing device based on wavelet multi-scale decomposition, comprising:

[0055] a multi-scale frequency decomposition module, configured to obtain a low-frequency subband and three high-frequency subbands by performing three-layer wavelet decomposition on the original foggy remote sensing image;

[0056] a high-frequency detail dynamic enhancement module, configured to perform nonlinear enhancement on the LH subband, the HL subband and the HH subband in the three high-frequency subbands respectively, to obtain three enhanced high-frequency subbands;

[0057] a low-frequency perception-guided filtering module, configured to synthesize a guided image according to the three enhanced high-frequency subbands, and perform guided filtering on the low-frequency subband according to the guided image, to obtain an enhanced low-frequency subband;

[0058] a multi-scale context aggregation module, configured to fuse the enhanced low-frequency subband and the three enhanced high-frequency subbands by using inverse discrete wavelet transform, to obtain preliminary fusion features; and perform hole convolution, residual connection and SE module on the preliminary fusion features, to obtain fusion features;

[0059] a decoding reconstruction module, configured to generate a dehazed remote sensing image by using three decoding submodules based on the fusion features and the three enhanced high-frequency subbands; wherein each decoding submodule includes: performing transpose convolution and bilateral filtering on input features of the current decoding submodule, and fusing the input features with the enhanced high-frequency subband corresponding to the current decoding submodule, to obtain detail features of the current decoding submodule; performing adaptive sharpening kernel on the detail features of the current decoding submodule, to obtain output features of the current decoding submodule, and taking the output features as input features of the next decoding submodule; and the input features of the third decoding submodule are the fusion features.

[0060] The above technical solutions of the present application have the following beneficial effects compared with the prior art:

[0061] The wavelet multi-scale decomposition-based remote sensing image defogging method provided by the application first separates the original foggy remote sensing image into a low-frequency subband and a high-frequency subband through multi-scale wavelet frequency decomposition, can effectively separate fog and haze interference and image details, so that the high-frequency details of the image are completely retained in the subsequent defogging process; then, a nonlinear dynamic enhancement function is used to strengthen the details of the high-frequency subband, the low-frequency subband is subjected to multi-scale guided filtering using the enhanced high-frequency subband, the fog component in the image is suppressed, so that excellent defogging effect is achieved in light / heavy fog scenes, and the definition and visual effect of the image are improved; further, the feature information at different scales is integrated through residual hole convolution and an SE module, the perception ability of the global structure and local details is strengthened, and fused features are obtained, so that the edge loss problem in the downsampling process can be avoided; finally, the fused features are decoded and reconstructed, and a high-resolution defogging remote sensing image is obtained. The application not only effectively improves the definition, structural integrity and application adaptability of the recovered remote sensing image under light to heavy fog conditions, but also reduces the operation consumption and improves the algorithm execution speed, so that the application can normally operate even in the case of limited resources, and provides a reliable data basis for remote sensing interpretation, target recognition and environmental monitoring tasks. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to make the content of the application more easily understood, the application will be further described in detail below according to specific embodiments of the application and in conjunction with the drawings, in which:

[0063] Figure 1 is a flowchart of the wavelet multi-scale decomposition-based remote sensing image defogging method provided by the application;

[0064] Figure 2 is a schematic diagram of wavelet decomposition of the original foggy remote sensing image;

[0065] Figure 3 is a flowchart of the multi-scale context aggregation module;

[0066] Figure 4 is a defogging effect visualization diagram of the method provided by the application on a public data set, in which Figure 4 (a) in the column is the original foggy remote sensing image, Figure 4 (b) in the column is the labeled clear remote sensing image, Figure 4 (c) in the column is the defogging remote sensing image. DETAILED DESCRIPTION

[0067] The application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the application and implement it, but the embodiments are not limiting to the application.

[0068] Due to the influence of atmospheric scattering, cloud and fog blocking and other factors, remote sensing images often face the problem of quality degradation, resulting in loss of key information and decline of analysis accuracy. Remote sensing image defogging technology as the core link of remote sensing image quality improvement and information reconstruction has irreplaceable engineering value in the fields of surface dynamic monitoring, disaster emergency response, ecological environment assessment, precision agriculture management, urban planning and resource exploration. By eliminating the degradation interference and restoring the image details, high-quality data basis is provided for subsequent interpretation and application.

[0069] In the field of computer vision, atmospheric scattering model is often used to describe the formation of degraded images, and its mathematical expression is:

[0070] ;

[0071] Among them, is the degraded image, is the clear image, is the global atmospheric light value, is the medium transmittance; from the known , the purpose of defogging is to recover , A and t.

[0072] The first term of the equation is called direct attenuation, which describes the scene radiance and its attenuation in the medium; the second term of the equation is called atmospheric light, which comes from previously scattered light and causes scene color shift. When the atmosphere is uniform, the medium transmittance t can be expressed as:

[0073] ;

[0074] Among them, is the total scattering coefficient, is the scene depth, which means that the scene brightness presents exponential decay with depth.

[0075] The remote sensing image defogging task aims to restore image quality, enhance features and improve the accuracy of subsequent application, and has irreplaceable core value in the fields of ecological environment monitoring, disaster emergency response and military reconnaissance.

[0076] Referring to the method for removing fog from a remote sensing image based on wavelet multi-scale decomposition provided by the present application, the method comprises the following steps: Figure 1 S1: after three-layer wavelet decomposition of the original foggy remote sensing image, a low-frequency subband and three high-frequency subbands are obtained.

[0077]

[0078] ​The object of the present application is to create a remote sensing image defogging algorithm enhanced by wavelet features, to achieve the best coordination between defogging effect and detail preservation by means of the multi-scale decomposition function of wavelet transform. Wavelet transform is a kind of time-frequency local analysis method, which can divide the image into low-frequency and high-frequency subbands. The low-frequency subband represents the general structure of the image, and the high-frequency subband corresponds to the details of the image. After strengthening the high-frequency features in the wavelet domain and suppressing the low-frequency noise, the original image details and color properties can be preserved to the maximum extent while removing the haze interference. The present application not only maintains the original advantages of multi-scale decomposition of wavelet, but also improves the image clarity and resolution through the feature enhancement link.

[0079] The present application proposes a multi-scale frequency decomposition module (MFD): inputting the original foggy remote sensing image into the multi-scale frequency decomposition module, and obtaining the low-frequency subband and three-layer high-frequency subband after three-layer wavelet decomposition of the original foggy remote sensing image. The module fully utilizes the localization advantages of wavelet transform in time domain and frequency domain, and effectively divides the remote sensing image into low-frequency and high-frequency subbands through multi-level decomposition. Among them, the low-frequency subband carries the structure and brightness information of the image, and the high-frequency subband contains edge, texture and other detail features.

[0080] Preferably, the present application adopts Daubechies-4 (db4) wavelet basis function for discrete wavelet transform of the original foggy remote sensing image, has tight support, high vanishing moments and good smoothness, and is suitable for preserving the spatial frequency variation of complex ground objects in remote sensing images.

[0081] Daubechies wavelet is a kind of orthogonal wavelet family with tight support, orthogonality and strong resolution capability, and db4 is one of the shorter length (4th order filter coefficient) members, which has the following obvious advantages:

[0082] (1) Good time-frequency localization characteristics: db4 wavelet has tight support characteristics in time domain and frequency domain, which can better preserve the local detail information of the image, especially suitable for processing local dense fog, cloud layer, edge structure and other detail areas in remote sensing images.

[0083] (2) Strong edge preservation ability: db4 wavelet has high vanishing moments, which is more helpful for data compression and noise elimination, can effectively represent the smooth area and abrupt boundary of the image, capture edge structure and texture details under multi-scale, and facilitate structure recovery in subsequent defogging enhancement.

[0084] (3) Balanced spatial and frequency resolution: Compared with the shorter basis function such as Haar wavelet, db4 can provide smoother filter response, avoiding the phenomenon of excessive frequency leakage, and is suitable for processing the rich low-frequency structure information in remote sensing images.

[0085] (4) Low computational complexity, suitable for engineering implementation: The filter length of db4 is moderate (supporting length of 8), which not only ensures the analysis accuracy, but also takes into account the computational efficiency, and is suitable for the implementation requirements of real-time or semi-real-time remote sensing image processing system.

[0086] In general, choosing db4 wavelet as the basic function of image wavelet decomposition can stably extract the multi-scale features of the image on the premise of maintaining the structure and detail information, thereby providing more accurate feature support for subsequent guided enhancement and image reconstruction.

[0087] Wavelet decomposition extracts the spatial frequency of the image through Daubechies-4 wavelet basis function, and decomposes the image into low-frequency sub-band (LL) and high-frequency sub-band (LH, HL, HH). The naming rule of the four sub-bands is the combination of "row frequency-column frequency". Among them: L (low): low-frequency component (corresponding to the area with low spatial frequency); H (high): high-frequency component (corresponding to the area with high spatial frequency); The first letter represents the frequency component in the row direction (horizontal direction); The second letter represents the frequency component in the column direction (vertical direction). Therefore, LL is: row low frequency+column low frequency, which represents the global low frequency and describes the main structure and outline of the image; LH is: row low frequency+column high frequency, which represents the horizontal low frequency and vertical high frequency, and describes the vertical edges / textures of the image; HL is: row high frequency+column low frequency, which represents the horizontal high frequency and vertical low frequency, and describes the horizontal edges / textures of the image; HH is: row high frequency+column high frequency, which represents the global high frequency and describes the details and noise of the image.

[0088] For the original foggy remote sensing image , the image is decomposed into low-frequency sub-band (LL) and high-frequency sub-band (LH, HL, HH) through the MFD module. These components are used to extract image information at different scales, where the low-frequency component contains the global information of the image, and the high-frequency component retains the detail and edge features of the image. This process is completed through successive wavelet decomposition, which uses low-pass filter h[n-2k] and high-pass filter g[n-2k] in the row and column directions respectively, and performs downsampling operation after each filtering. Finally, the image is divided into four sub-bands, and the decomposed image can be effectively used for defogging processing. The decomposition process uses Daubechies-4 (db4) wavelet basis function as the core basis function for multi-scale frequency extraction.

[0089] Specifically, referring to Figure 2As shown, the original foggy remote sensing image is decomposed into low-frequency subband and three high-frequency subbands by using Daubechies-4 wavelet basis function, including:

[0090] S11: Discrete Wavelet Transform (DWT) is performed on the original foggy remote sensing image using db4 wavelet basis function to obtain the first layer wavelet decomposition result, denoted as:

[0091] ;

[0092] wherein, is the first layer low-frequency subband, representing the main structural information of the image; are the first layer high-frequency subbands, respectively representing the horizontal, vertical and diagonal line direction detail information in the image, is the discrete wavelet transform.

[0093] The first layer wavelet decomposition is specifically expanded as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] wherein, and are low-pass filter and high-pass filter respectively, is the radius of the filter; "*" represents convolution operation; represents the original input signal, i.e. the original foggy remote sensing image, represents the discrete position or index of data; and represent down-sampling operation row by row and column by column respectively, and the embodiment samples with fixed step length 2, only retains pixel values at even or odd positions to realize reduction of image resolution and retain specific frequency components.

[0099] S12: After completing the first layer wavelet decomposition, the first layer low-frequency subband is decomposed into the second layer wavelet, and the second layer low-frequency subband is decomposed into the third layer wavelet to obtain low-frequency subband and high-frequency subband with finer granularity. The second layer and third layer wavelet decomposition are denoted as:

[0100] ;

[0101] wherein, and​ are the jth and (j-1)th low frequency subbands, respectively; are the jth high frequency subbands, respectively; .

[0102] The second and third layer wavelet decompositions are specifically expanded as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] wherein, represents the (j-1)th low frequency subband.

[0108] After the original foggy remote sensing image is decomposed by three layers of wavelets, low frequency subbands and three layers of high frequency subbands are obtained, which can be represented by an MFD module as follows:

[0109] ;

[0110] wherein, is the original foggy remote sensing image, is a multi-scale frequency decomposition module, is the third low frequency subband, are the jth high frequency subbands, respectively.

[0111] will be used in the following to represent the low frequency subband, to represent the high frequency subband.

[0112] The advantages of the MFD module in remote sensing image defogging are embodied in two aspects: on the one hand, the low frequency subband can be used to model the haze component caused by atmospheric scattering and is inhibited in subsequent guided filtering; on the other hand, the high frequency subband can realize fine detail restoration of direction perception through subband separation enhancement strategy, significantly improving the image edge definition and texture fidelity. Compared with traditional single-scale methods, the MFD module realizes the frequency domain decoupling of image structure and fog interference, lays a stable foundation for subsequent feature enhancement and defogging reconstruction, and exhibits superior clear recovery ability and detail preservation performance under light / heavy fog conditions.

[0113] S2: Perform nonlinear enhancement on the LH subband, HL subband and HH subband in the three layers of high frequency subbands, respectively, to obtain three layers of enhanced high frequency subbands.

[0114] The application proposes a high-frequency detail dynamic enhancement module (HDDE): three-layer high-frequency subbands are input into the high-frequency detail dynamic enhancement module, and the LH subband, HL subband and HH subband in the high-frequency subband are respectively subjected to nonlinear enhancement, so as to obtain three-layer enhanced high-frequency subbands.

[0115] Specifically, the horizontal direction LH subband in the three-layer high-frequency subband is subjected to nonlinear enhancement by using an S-shaped inhibition enhancement function, so as to inhibit the response to small amplitude coefficients and prevent flat areas from being enlarged; and large coefficients are gradually enhanced, which is beneficial to retaining clear horizontal edges. The formula is as follows:

[0116] ;

[0117] wherein, is the enhanced LH subband, is the LH subband, is an enhancement ratio of the LH subband set according to local texture intensity, which is used for controlling the overall enhancement ratio of the LH subband in the horizontal direction; is a position parameter of the LH subband set according to the haze masking degree, which is used for controlling the "saturation position" of the subband enhancement function in the horizontal direction.

[0118] The HL subband in the three-layer high-frequency subband is subjected to nonlinear enhancement by using an exponential response enhancement function, which has a fast nonlinear response capability and can significantly enhance the vertical strong edge region of buildings and the like, while having a small response to noise. The formula is as follows:

[0119] ;

[0120] wherein, is the enhanced HL subband, is the HL subband, is an enhancement ratio of the HL subband, which is used for controlling the overall enhancement ratio of the HL subband in the vertical direction; is an enhancement response speed of the HL subband set according to local texture intensity, and the greater the value, the more rapid the function responds under small input, which is suitable for fast edge enhancement.

[0121] The HH subband in the three-layer high-frequency subband is subjected to nonlinear enhancement by using a double-gated nonlinear enhancement function, which controls the enhancement response to weak texture and large-scale diagonal edges by introducing two gating functions with different sensitivities, so as to adapt to complex texture areas and mixed direction signals. The formula is as follows:

[0122] ;

[0123] wherein, ​for the enhanced HH subband, for the HH subband, for the enhanced HH subband, for controlling the overall enhancement ratio of the HH subband in the diagonal direction; and are both gating parameters, adjusted depending on the local fog concentration, is a smaller value, used to control the weak texture enhancement, with a wider activation range, is a larger value, used to control the strong structure edge activation.

[0124] In general, the is the "enhancement magnification"; is the "when to start enhancement" controller; mainly adjusts the "enhancement speed"; is the "gating response speed". Multiple scale responses can be combined to achieve image content adaptive adjustment, improving applicability and robustness.

[0125] The final enhanced high-frequency subband representation is .

[0126] The HDDE module aims to enhance the edge structure and texture details in remote sensing images that are easily obscured by fog by implementing direction-aware nonlinear enhancement on three high-frequency subbands (LH, HL, and HH). The module designs differentiated adaptive enhancement functions based on the frequency characteristics and feature distribution of each directional subband to strengthen the detail response in the horizontal, vertical, and diagonal directions. By introducing S-type suppression functions, exponential response enhancement functions, and double-gating 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, the module adaptively adjusts the enhancement intensity based on prior information such as local texture intensity and fog obscuration, significantly improving image clarity and structural integrity in different concentration fog scenes. The module not only enhances weak but crucial edge features in remote sensing images but also provides more abundant and reliable detail support for subsequent low-frequency guided filtering and fusion reconstruction, thereby improving the overall visual quality and information fidelity of the dehazed image.

[0127] S3: Synthesize the guide image according to the three enhanced high-frequency subbands, and then perform guided filtering on the low-frequency subband according to the guide image to obtain the enhanced low-frequency subband.

[0128] The application proposes a low-frequency-aware guided filter module (LGF): input the low-frequency subband and the three-layer enhanced high-frequency subband into the low-frequency-aware guided filter module, synthesize a guided image according to the three-layer enhanced high-frequency subband, and then perform guided filtering on the low-frequency subband according to the guided image to obtain an enhanced low-frequency subband. The module is designed for the feature that the fog component in a remote sensing image is mainly concentrated in a low-frequency region, and through the introduction of high-frequency details as a guide, effective suppression of the fog in the low-frequency subband and explicit recovery of structural information are realized.

[0129] First, the three-layer enhanced high-frequency subband is fused as a detail map as the guided image of the original low-frequency subband , which is expressed by the formula:

[0130] ;

[0131] Among them, is the guided image, is the jth layer of enhanced high-frequency subband.

[0132] The target of guided filtering is to solve the filter output , and the essence is to establish the local linear relationship between and , which specifically includes:

[0133] For each pixel , in the window centered on , it is assumed that the output and the guide satisfy the linear relationship:

[0134] ;

[0135] Among them, is the enhanced low-frequency subband, i is the pixel index, is the first linear coefficient in the window , is the second linear coefficient in the window , is the window.

[0136] Solve and by minimizing the cost function:

[0137] ;

[0138] Among them, is the low-frequency subband, For low-frequency subband in pixels The value; This is a regularization parameter used to prevent the image from becoming too smooth.

[0139] Solving the above cost function yields and :

[0140] ;

[0141] 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.

[0142] 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:

[0143] 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:

[0144] ;

[0145] 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, .

[0146] S32: Calculate the adaptive weights of the filtering results for each window, using the following formula:

[0147] ;

[0148] 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, is the local haze density of the mth window, is the low-frequency subband;

[0149] scale-sensitive parameter wherein is the window radius of the mth window, is a constant.

[0150] S33: obtaining the enhanced low-frequency subband according to the filtering results of each window and the adaptive weight, the formula being:

[0151] ;

[0152] wherein, is the enhanced low-frequency subband.

[0153] Specifically, the calculation formula of the first linear coefficient and the second linear coefficient of the mth window is:

[0154] ;

[0155] ;

[0156] wherein, is the first linear coefficient of the mth window, is the second linear coefficient of the mth window, is the mth window, is the number of pixels in the mth window, i is the pixel index, is the guide image, is the low-frequency subband, is the mean value of the low-frequency subband in the mth window, and are respectively the mean value and the variance of the guide image in the mth window, is a regularization parameter.

[0157] Compared with the single-scale guided filtering, i.e. using a fixed window radius to perform filtering, the multi-scale guided filtering proposed in the embodiment has the following significant advantages:

[0158] (1) Strong structure preservation: by performing filtering under multiple receptive fields, the overall contour and edge structure of the image can be more fully extracted, effectively alleviating the structure breaking problem caused by haze blocking at small scales;

[0159] (2) Good adaptability: the adaptive weight fusion function is introduced to adaptively adjust the multi-scale response according to the local brightness difference at each pixel, effectively avoiding over-smoothing or over-enhancement caused by improper scale selection.

[0160] (3) Strong noise suppression capability: This method can control the propagation of high-frequency noise while preserving the structure perception ability of the image, avoiding the introduction of texture artifacts when enhancing edges by traditional guided filtering;

[0161] (4) Strong coupling with subsequent modules: The output weighted low-frequency estimation result can better integrate low-frequency brightness information and high-frequency structure perception information, providing a clearer, more natural, and structure-accurate image foundation for subsequent operations, and participating in the overall reconstruction of the defogging image as a structure guiding signal, improving the overall restoration quality.

[0162] The LGF module performs local linear guided filtering at multiple scales, adaptively adjusting the filtering weights to adapt to the differences in fog concentration in different regions, thereby effectively removing the blur and occlusion in the low-frequency component while preserving the ground object contour and hierarchical structure. Compared with traditional single-scale filtering methods, the LGF module has stronger structure preservation ability, adaptability, and noise suppression performance, and can avoid detail loss and edge blur caused by excessive smoothing. In the entire defogging method, the LGF module is located after the high-frequency detail enhancement and before the feature fusion, and undertakes the core task of bridging fog suppression and structure enhancement: on the one hand, it removes the remaining fog through accurate modeling of low-frequency brightness information, and on the other hand, it provides a clearer and higher-contrast structure foundation for subsequent feature fusion, significantly improving the clarity and naturalness of the final defogging image. It is the key link to realize the collaborative optimization of frequency domain decoupling and spatial reconstruction.

[0163] In summary, by introducing the LGF module, the present application effectively solves the problems of structure blur, edge blur, and loss of hierarchy caused by local fog or low-contrast regions in the traditional low-frequency component processing process. The module performs guided filtering operations at multiple scale windows and adaptively fuses the outputs of each scale based on local fog differences, thereby achieving enhanced perception and detail recovery of low-frequency structure information in the image.

[0164] S4: Fuse the enhanced low-frequency subband and the three enhanced high-frequency subbands using discrete wavelet inverse transform to obtain preliminary fusion features; pass the preliminary fusion features through a hole convolution, residual connection, and SE module to obtain fusion features.

[0165] 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.

[0166] 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.

[0167] Reference Figure 3 As shown, the specific steps of S4 include:

[0168] 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.

[0169] 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:

[0170] ;

[0171] in, As a preliminary fusion feature, This is the inverse discrete wavelet transform.

[0172] 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:

[0173] ;

[0174] 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.

[0175] 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.

[0176] 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:

[0177] ;

[0178] in, For the first Each residual dilated convolution outputs a feature.

[0179] 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.

[0180] 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.

[0181] 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:

[0182] 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:

[0183] ;

[0184] in, The first channel vector an element, and are the height and width of the residual dilated convolution output feature, respectively;

[0185] According to , the channel vector is obtained.

[0186] S44-2: Channel Excitation: input the channel vector into a two-layer fully connected network to model the nonlinear relationship between channels, calculate the weight of each residual dilated convolution output feature, and generate normalized weight , where is the weight of the th residual dilated convolution output feature. The specific formula is:

[0187] ;

[0188] where, is the normalized weight, is the weight matrix of the two-layer fully connected, is the compression ratio (default set to 2 or 4); is the Sigmoid function, used to normalize the output to the interval; is the ReLU activation function.

[0189] S44-3: Channel Reweight: according to the weight , the residual dilated convolution output feature is weighted and summed to obtain the fusion feature, realizing channel reweighting and weighted summation, and the formula is:

[0190] ;

[0191] where, is the fusion feature.

[0192] By introducing the channel attention mechanism of the SE module, the response ability of key features can be effectively improved, the directional perception and structure preservation ability in the defogging process can be enhanced, and the enhanced fusion feature , which is the fusion of high-frequency detailed information and low-frequency structural information under multiple scales, is obtained as the input of the subsequent decoding and reconstruction stage.

[0193] The MCA module first captures multi-scale spatial semantic information of the image under different receptive fields using a hollow convolution, which is particularly suitable for complex scenes in remote sensing images with varying scales of ground objects and uneven distribution of fog; then, an SE channel attention mechanism is introduced to adaptively adjust the importance weight of each subband, highlighting key structures and edge details, and effectively alleviating the texture loss problem caused by downsampling. In the entire dehazing system, the MCA module serves as the core fusion unit connecting the frequency domain enhancement (HDDE, LGF) and the decoding reconstruction module, not only integrating multi-source information to form a unified feature representation, but also improving the compatibility of global consistency and local fineness, ensuring that the final restored image has high structure restoration, clarity and naturalness. It is the key hub link for realizing multi-scale information collaborative optimization and full-chain enhancement reconstruction.

[0194] S5: Based on the fused features and the high-frequency subbands enhanced by three layers, a dehazed remote sensing image is generated using three decoding submodules; wherein each decoding submodule includes: the input features of the current layer decoding submodule are fused with the high-frequency subbands enhanced by the current layer after being processed by transposed convolution and bilateral filtering, to obtain the detail features of the current layer decoding submodule; the detail 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 used as the input features of the next layer decoding submodule; the input features of the third layer decoding submodule are the fused features.

[0195] The decoding reconstruction module for detail recovery is the final output link of the method, which is mainly responsible for restoring the fused features after multi-scale fusion into a high-resolution clear remote sensing image, and enhancing edges and textures through detail injection and adaptive sharpening.

[0196] Specifically, the fused features are input into the third layer decoding submodule, fused with the high-frequency subbands enhanced by the third layer, to obtain the third layer decoding features; the third layer decoding features are sequentially input into the second layer decoding submodule and the first layer decoding submodule, respectively fused with the high-frequency subbands enhanced by the second layer and the first layer, to obtain the dehazed remote sensing image. The specific steps of each decoding submodule include:

[0197] S51: The input features of the current layer decoding submodule are processed by transposed convolution to obtain up-sampling features, which are represented by the formula:

[0198] ;

[0199] wherein, is the input features of the jth layer decoding submodule, denotes the transposed convolution operation, is the up-sampling features of the jth layer decoding submodule; is the index of the level of the wavelet decomposition and decoding submodule.

[0200] 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.

[0201] 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, thereby generating a dehazed high-resolution remote sensing image.

[0202] 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:

[0203] ;

[0204] 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.

[0205] 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.

[0206] 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:

[0207] ;

[0208] 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... to balance the contribution of fused features and high-frequency details; for inverse discrete wavelet transform, for the jth layer enhanced high-frequency subband.

[0209] S54: To further enhance the image edges, an adaptive sharpening kernel is designed based on the local gradient magnitude, and the detail features of the current layer decoding sub-module are passed through the adaptive sharpening kernel to obtain the output features of the current layer decoding sub-module, which are used as the input features of the next layer decoding sub-module, including:

[0210] The gradient magnitude of the detail features of the current layer decoding sub-module is calculated, and the formula is:

[0211] ;

[0212] wherein, is the detail feature of the jth layer decoding sub-module, and are the gradient magnitudes of the x-direction and y-direction of the detail feature of the jth layer decoding sub-module, and are the convolution kernels of the x-direction and y-direction of the operator, is the gradient magnitude of the detail feature of the jth layer decoding sub-module;

[0213] The sharpening intensity of the current layer decoding sub-module is calculated according to the gradient magnitude of the detail feature of the current layer decoding sub-module, and the formula is:

[0214] ;

[0215] wherein, is the sharpening intensity of the jth layer decoding sub-module, is a scaling factor for controlling the range of sharpening intensity; is a Sigmoid function for mapping the gradient to the interval [0, 1]; is a strong / weak edge threshold for distinguishing strong / weak edge regions;

[0216] The detail features are sharpened according to the sharpening intensity of the current layer decoding sub-module to obtain the output features of the current layer decoding sub-module, which are used as the input features of the next layer decoding sub-module, and the formula is:

[0217] ;

[0218] wherein, is the input feature of the j-1th layer decoding sub-module, is a Gaussian blur (kernel size 3x3, standard deviation 1); for the Laplacian approximation.

[0219] the output feature of the last decoding sub-module for the defogged remote sensing image output by the decoding reconstruction module, the formula is:

[0220] ;

[0221] wherein, the defogged remote sensing image.

[0222] The decoding reconstruction module realizes step-by-step up-sampling through multi-level transposed convolution, and introduces bilateral filtering to eliminate artifacts that may be introduced in the up-sampling process, while injecting enhanced high-frequency detail subbands through a skip connection to further restore the edges and textures lost due to down-sampling or haze blocking. In order to further strengthen the image sharpness and level of detail, the module also designs an adaptive sharpening mechanism combining local gradient information to achieve precise enhancement of strong edge regions. The advantage of this module is that on the one hand it ensures the clarity and structural integrity of the image reconstruction process, avoiding the common problems of blurring and distortion in traditional decoding; on the other hand, through close cooperation with the frequency domain enhancement module (HDDE, LGF) and the multi-scale fusion module (MCA), it realizes the whole-chain collaborative optimization from frequency domain decomposition to spatial restoration. In the entire system, the decoding reconstruction module not only undertakes the final task of image generation, but also effectively converts the results of the previous feature mining and enhancement into a key node for improving the quality of visual images, directly determining the practicality and visual performance of the final remote sensing image.

[0223] In order to verify the effectiveness of the method of the present application, the present embodiment also selects the public dataset SateHaze1k of remote sensing images as the experimental data source. The dataset consists of 1200 pairs of clear synthetic aperture radar (SAR), foggy visible light remote sensing images and corresponding ground true images, according to the different fog concentration, i.e. the degradation degree of the remote sensing image, it is divided into three sets: thin fog (StateHaze1k-TN), medium fog (StateHaze1k-M) and thick fog (StateHaze1k-TK), each containing 400 images. In the synthetic image covering thin fog, the fog mask is only the fog extracted from the original true cloudy image. For medium fog images, the sample is the superposition of fog and medium concentration fog. And for thick fog images, the transmission map is selected from thick fog.

[0224] The size of the input original foggy remote sensing image is set to 512x512: the colors representing different land cover categories in the RGB label are mapped to corresponding gray values, and the image and label are cropped to meet the input size requirements of the model. After cropping, the image size is uniformly changed to 512x512 pixels, and then the cropped image is input into the deep network for training.

[0225] From a total of 3 groups of 400 sample images, 350 sample images are randomly selected from each group for training the deep network model, and 30 samples are used for verification data set. After the end of training, all the remaining sample pictures are input into the deep network model for testing, and finally the dehazing results of the test samples are obtained. Peak signal-to-noise ratio and structural similarity are selected to evaluate the classification results.

[0226] The dehazing results obtained by the remote sensing image dehazing method, the traditional method and the deep learning method proposed in the application are shown in Table 1.

[0227] Table 1, comparison of experimental results

[0228]

[0229] As can be seen from the table, in the light haze scene, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the method of the application are 32.35dB and 0.925 respectively, which are 3.5dB and 0.065 higher than those of the traditional method (28.85dB, 0.86) respectively, and 1.35dB and 0.015 higher than those of the deep learning method (31dB, 0.91) respectively, indicating that the method has obvious advantages in detail recovery and structure preservation.

[0230] In the heavy haze scene, the PSNR and SSIM of the method of the application are 30.89dB and 0.968 respectively, which are more significantly improved than those of the traditional method (26.95dB, 0.825), and are increased by 3.94dB and 0.143 respectively; compared with the deep learning method (28.9dB, 0.88), it is also increased by 1.99dB and 0.088 respectively. This fully proves that the application can still maintain good image reconstruction ability and structure fidelity under high haze concentration conditions.

[0231] Figure 4 The dehazing effect of the method of the application on the public data set is visualized in the figure, wherein Figure 4 (a) in the figure is the original hazy remote sensing image, Figure 4 (b) in the figure is the clear remote sensing image, Figure 4 (c) in the figure is the dehazed remote sensing image.

[0232] As can be seen, the wavelet feature enhanced remote sensing image dehazing algorithm proposed in the application has achieved good results in both light haze scene and heavy haze scene. Compared with the traditional method and the deep learning method, the application has greatly improved the PSNR and SSIM indicators, indicating that the algorithm has obvious advantages in improving image clarity and preserving image details.

[0233] Therefore, the method of the present application realizes the quantitative performance better than the existing method under different haze intensities, especially shows stronger robustness and recovery ability in heavy haze scenes, verifying the significant effectiveness of the designed wavelet multi-scale enhancement and detail guidance mechanism in the remote sensing image defogging task.

[0234] In summary, the remote sensing image defogging method based on wavelet multi-scale decomposition provided by the present application firstly separates the original foggy remote sensing image into a low-frequency subband and a high-frequency subband through multi-scale wavelet frequency decomposition, which can effectively separate the haze interference and image details, so as to completely retain the high-frequency details of the image in the subsequent defogging process; then a nonlinear dynamic enhancement function is used to strengthen the details of the high-frequency subband, and the enhanced high-frequency subband is used to guide the filtering of the low-frequency subband, so as to suppress the fog component in the image and achieve excellent defogging effect in light / heavy haze scenes, and improve the clarity and visual effect of the image; further, the feature information at different scales is integrated through residual hole convolution and SE module to strengthen the perception ability of global structure and local details, and the fusion features are obtained, which can avoid the edge loss problem in the downsampling process; finally, the fusion features are decoded and reconstructed to restore the high-resolution defogged remote sensing image. The present application not only effectively improves the clarity, structural integrity and application adaptability of the restored remote sensing image under light to heavy haze conditions, but also reduces the operation consumption and improves the algorithm execution speed, so that it can normally operate even in the case of limited resources, and provides a reliable data basis for remote sensing interpretation, target recognition and environmental monitoring tasks.

[0235] Based on the remote sensing image defogging method based on wavelet multi-scale decomposition, the present application further provides a remote sensing image defogging device based on wavelet multi-scale decomposition, comprising:

[0236] A multi-scale frequency decomposition module is configured to perform three-layer wavelet decomposition on the original foggy remote sensing image to obtain a low-frequency subband and three high-frequency subbands.

[0237] A high-frequency detail dynamic enhancement module is configured to perform nonlinear enhancement on the LH subband, HL subband and HH subband in the three high-frequency subbands respectively to obtain three enhanced high-frequency subbands.

[0238] A low-frequency perception guided filtering module is configured to synthesize a guided image according to the three enhanced high-frequency subbands, and then perform guided filtering on the low-frequency subband according to the guided image to obtain an enhanced low-frequency subband.

[0239] A multi-scale context aggregation module is configured to fuse the enhanced low-frequency subband and the three enhanced high-frequency subbands by using inverse discrete wavelet transform to obtain preliminary fusion features; and the preliminary fusion features are subjected to hole convolution, residual connection and SE module to obtain fusion features.

[0240] The decoding reconstruction module is configured to generate the dehazed remote sensing image by using the three-layer decoding sub-modules based on the fused features and the three-layer enhanced high-frequency subbands; each of the decoding sub-modules comprises: fusing the input features of the current decoding sub-module after being subjected to transposed convolution and bilateral filtering with the high-frequency subband corresponding to the current layer after enhancement to obtain the detail features of the current decoding sub-module; subjecting the detail features of the current decoding sub-module to an adaptive sharpening kernel to obtain the output features of the current decoding sub-module and taking the output features as the input features of the next decoding sub-module; and the input features of the third decoding sub-module are the fused features.

[0241] Obviously, the above embodiments are only examples for clearly illustrating the present application, and are not intended to limit the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments are not required to be exhausted, and the obvious changes or variations derived therefrom are still within the protection scope of the present application.

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 layer of the decoding submodule includes: after transposing the input features of the current layer's decoding submodule through convolution and bilateral filtering, fusing them with the enhanced high-frequency subband corresponding to the current layer to obtain the detailed features of the current layer's decoding submodule; 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 dehazed remote sensing image output by the decoding and reconstruction module; The input features of the third-layer decoding submodule are 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 for 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 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 layer of the decoding submodule includes: after transposing the input features of the current layer's decoding submodule through convolution and bilateral filtering, fusing them with the enhanced high-frequency subband corresponding to the current layer to obtain the detailed features of the current layer's decoding submodule; 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 dehazed remote sensing image output by the decoding and reconstruction module; The input features of the third-layer decoding submodule are fused features.

Images