A remote sensing image defogging method, system, medium and device

Abstract

The application discloses a kind of remote sensing image defogging method, system, medium and equipment, it is related to image defogging technical field, including obtaining the original fog image of the region to be measured;Through lightweight deep learning model L-RSDformer, original fog image is preliminarily defogged, and preliminary defogging result is determined;Multi-stage optimization mathematical model PMG-Defog is constructed;Preliminary defogging result and original fog image are as the input of multi-stage optimization mathematical model PMG-Defog, and through transmission map refinement mathematical model and atmospheric light self-adaptive estimation algorithm optimization process, optimal fog attenuation graph and optimal atmospheric light value are obtained;According to optimal fog attenuation graph, determine refined transmission map;Refined transmission map and the atmospheric light graph obtained are input to re-formalized atmospheric scattering model and are inverted, and final defogging image is obtained.

Description

Technical Field

[0001] This invention relates to the field of image dehazing technology, and in particular to a method, system, medium, and device for dehazing remote sensing images. Background Technology

[0002] Remote sensing technology is widely used in military reconnaissance, environmental monitoring, and land spatial planning. However, haze can cause remote sensing images to suffer from reduced contrast, blurred details, and dull colors, which seriously affects the accuracy of subsequent computer vision tasks such as target detection and image classification.

[0003] Existing dehazing methods, such as pure data-driven deep learning models, can efficiently remove haze from the main subject and restore details. However, due to the lack of explicit physical consistency constraints, problems such as unnatural brightness and color transitions in areas of abrupt changes in depth of field, and slight noise or uneven textures in large uniform areas may occur, which makes it impossible to guarantee the dehazing effect. Summary of the Invention

[0004] This invention provides a method, system, medium, and device for dehazing remote sensing images to solve the aforementioned problems in the prior art, namely, how to improve the dehazing effect of remote sensing images in the prior art. This invention provides a method for dehazing remote sensing images, which includes: Acquire the original haze image of the area to be tested; A pre-trained lightweight deep learning model, L-RSDformer, is used to perform preliminary dehazing on the original haze image to determine the preliminary dehazing result. The L-RSDformer model comprises an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer, connected sequentially. The input layer maps the original haze image to high-dimensional initial features. The multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features. The DuRB performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features. The multi-scale decoding layer upsamples the global features to determine enhanced features. The output layer maps the enhanced features to the preliminary dehazing result. The preliminary dehazing results and the original haze image are used as inputs to the constructed multi-stage optimization mathematical model PMG-Defog. Based on the preliminary dehazing results and the original haze image, the initial haze attenuation map and the initial atmospheric light map are determined. By fixing the initial atmospheric light map, the energy function of the transmission map is minimized to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, the adaptive estimation energy function of atmospheric light is minimized to determine the atmospheric light value for the current iteration. The iterative process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing a comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model (ASM) is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazing image.

[0005] Optionally, obtaining the re-formulated atmospheric scattering model specifically includes: The comprehensive haze impact factor ρ(x) is obtained using the following formula: ρ(x) = β(x)・d(x) + δ(x); Where β(x) is the spatially varying scattering coefficient, d(x) is the scene depth, and δ(x) is the multiple scattering compensation term. Based on the comprehensive haze impact factor ρ(x), and combined with the spatially varying atmospheric light map A(x), the re-formulated atmospheric scattering model is obtained using the following formula: I(x)=J(x)・e (-[β(x)d (x)+δ(x)]) +A(x)(1-e (-[β(x)d(x)+δ(x)]) ); Where I(x) is the haze image, J(x) is the clear image, A(x) is the atmospheric light map, and the transmission map t(x) = e (-ρ(x)) .

[0006] Optionally, determining the refined transmittance map based on the optimal haze attenuation map specifically includes: Based on the optimal haze attenuation map, the refined transmittance map is determined using the following formula: T(x) = max(e (-D*(x)) ,t0); Where t0 is the lower limit of the transmission map, T(x) is the refined transmission map, and D * (x) is the optimal haze attenuation map.

[0007] Optionally, the step of inputting the refined transmission map and the optimized atmospheric light map into the reformulated atmospheric scattering model for inversion to obtain the final dehazed image specifically includes: J refined(x)=(I(x)-A*(x)(1-T(x))) / T(x); Among them, J refined (x) is the final dehazed image, A*(x) is the optimized atmospheric light map, and T(x) is the refined transmission map.

[0008] Optionally, the obtained smoothing constraint term and edge preservation term are weighted by introducing a regularization parameter, and the obtained data fidelity term, as well as the weighted smoothing constraint term and edge preservation term, are summed to determine the refined energy function of the transmission map.

[0009] Optionally, the acquisition of the atmospheric light adaptive estimation energy function specifically includes: The atmospheric light adaptive estimation energy function is obtained as follows: Where I(x) is the haze image and A(x) is the atmospheric light image. To smooth the intensity control parameters, For preliminary defogging results, This is a refined transmission image.

[0010] Optionally, the original haze image may be normalized before initial dehazing.

[0011] This invention provides a remote sensing image dehazing system, comprising: The acquisition module is used to acquire the original haze image of the area to be tested; The preliminary dehazing module is used to perform preliminary dehazing on the original haze image using a pre-trained lightweight deep learning model, L-RSDformer, to determine the preliminary dehazing result. The lightweight deep learning model L-RSDformer includes an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer connected in sequence. Specifically, the input layer maps the original haze image to high-dimensional initial features; the multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features; the dual residual bottleneck layer (DuRB) performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features; the multi-scale decoding layer upsamples the global features to determine enhanced features; and the output layer maps the enhanced features to the preliminary dehazing result. The dehazing module takes the preliminary dehazing results and the original haze image as input to the constructed multi-stage optimization mathematical model PMG-Defog. Based on the preliminary dehazing results and the original haze image, it determines the initial haze attenuation map and the initial atmospheric light map. By fixing the initial atmospheric light map, it minimizes the refined energy function of the transmission map to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, it minimizes the adaptive estimation energy function of atmospheric light to determine the atmospheric light value for the current iteration. The iteration process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing a comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model (ASM) is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazed image.

[0012] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described remote sensing image dehazing method.

[0013] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described remote sensing image dehazing method.

[0014] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention provides a remote sensing image dehazing method. This method, by fusing a lightweight deep learning model L-RSDformer with a physical model-guided multi-stage optimization mathematical model PMG-Defog, can determine a refined transmission map, resulting in a smoother and more natural dehazing result. This is particularly effective in large, uniform areas and object edges, improving visual quality and detail preservation, thus enhancing the dehazing effect. Furthermore, by using the preliminary dehazing result from L-RSDformer and the refined transmission map as priors, atmospheric light is estimated by solving an optimization problem constrained by the physical model. Accurate atmospheric light estimation helps restore the true color and brightness of the image, especially in distant and bright areas, improving the overall contrast and color saturation of the image. In addition, by inputting the refined transmission map and the acquired atmospheric light map into a reformed atmospheric scattering model for inversion, the reformed atmospheric scattering model effectively corrects the local biases of the pure data-driven model by modeling spatially varying atmospheric light and comprehensive haze influencing factors. This ensures that the brightness and color transitions in different depth areas conform to atmospheric perspective effects, thereby making the image dehazing effect more consistent with actual physical conditions. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0016] Figure 1 A flowchart of a remote sensing image dehazing method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the cascade fusion of the PMG-Defog model and the L-RSDformer, as provided in this embodiment of the invention. Figure 3 The iterative optimization convergence curve provided in the embodiments of the present invention; Figure 4 A schematic diagram of a computer device for a remote sensing image dehazing method provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0019] Example 1 Figure 1 This is a flowchart of a remote sensing image dehazing method provided in an embodiment of the present invention, such as... Figure 1 As shown in this embodiment, a method for dehazing remote sensing images includes: S1: Obtain the original haze image of the area to be tested.

[0020] For example, the acquired raw haze image can be an RGB three-channel image with a size of 512×512 pixels.

[0021] S2: The original haze image is initially dehazed using a pre-trained lightweight deep learning model, L-RSDformer, to determine the initial dehazing result. The lightweight deep learning model L-RSDformer includes an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer connected in sequence. The input layer maps the original haze image to high-dimensional initial features. The multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features. The dual residual bottleneck layer (DuRB) performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features. The multi-scale decoding layer upsamples the global features to determine enhanced features. The output layer maps the enhanced features to the initial dehazing result.

[0022] For example, the original haze remote sensing image I(x) is normalized to the [0,1] interval and input into the L-RSDformer model; coarse restoration stage: L-RSDformer extracts multi-scale frequency features through dynamic filters, and outputs the preliminary dehazing result after dynamic gating fusion, that is, the coarse restoration image J(x).

[0023] For example, the present invention can normalize the original haze remote sensing image I(x) to the [0,1] interval and input it into the L-RSDformer model; coarse recovery stage: L-RSDformer adopts an encoder-decoder U-shaped symmetrical structure, and the core modules include an input layer, a multi-scale coding layer (including a multi-depth convolution head transposed attention module, a gated depth convolution feedforward network module, a dual-frequency domain adaptive module, and a dual residual module), a multi-scale decoding layer (including a dynamic gated fusion module), and an output layer.

[0024] The input layer maps the 3-channel haze image into a high-dimensional initial feature F0 through two 3×3 convolutional layers, preserving the global structure and local texture of the remote sensing image. Multi-scale coding layer: The 4-level DFAB module generates spatial adaptive convolution kernels through dynamic filters, extracts multi-scale frequency features, and combines lightweight window attention with GDFN dynamic gated feedforward network to complete feature purification with low parameter quantity, and outputs multi-scale encoder features Fe1~Fe4. Dual Residual Bottleneck Layer (DuRB): The DuRB dual residual block performs global fusion and channel attention enhancement on the deepest layer features, outputting globally purified features Fb, which provides global context for the decoder; Multi-scale decoding layer: The decoder fuses high and low layer features of the encoder through 4-level upsampling, skip connections and dynamic gating fusion, and outputs the same resolution enhanced feature Fout; Output layer: The enhanced features are mapped to a 3-channel RGB image through convolutional layers and residual connections, and the final output is the preliminary dehazing result, i.e., the coarse restored image J(x).

[0025] Data flow between modules is achieved through feature map passing step by step and cross-layer skip connections. The multi-scale features extracted by DFAB are passed to the decoder after being gated by GDFN. The bottleneck layer ensures the consistency of global features, and finally provides reliable initial dehazing results and depth priors for subsequent PMG-Defog multi-stage optimization.

[0026] S3: Using the preliminary dehazing results and the original haze image as input to the constructed multi-stage optimization mathematical model PMG-Defog, the initial haze attenuation map and initial atmospheric light map are determined based on the preliminary dehazing results and the original haze image. By fixing the initial atmospheric light map, the energy function of the transmission map refinement is minimized to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, the atmospheric light adaptive estimation energy function is minimized to determine the atmospheric light value for the current iteration. The iteration process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing the comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model ASM is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazing image.

[0027] For example, this invention models the refinement of the haze attenuation map as an optimization problem based on a variational framework. Let the haze attenuation map to be refined be... , corresponding to the transmission diagram The present invention constructs the following energy function for minimization, as shown in formula (1):

[0028] Where E(D) is the total energy function to be minimized, and the optimal haze attenuation map D is obtained by solving its global minimum value. ∗ (x); The data fidelity term is a core component of the total energy function E(D), used to ensure the optimal haze attenuation map D after refinement. ∗ (x) A reliable initial estimate that does not deviate from the L-RSDformer output; To smooth out constraint terms and promote Continuity in space; This is an edge-preserving term designed to retain sharp transmission variations at object boundaries and abrupt changes in depth. and This is a regularization parameter used to balance the contributions of each term.

[0029] For example, the present invention uses data fidelity items Measure the fine-grained attenuation map Compared with the initial estimate Consistency between them. Considering that the confidence levels of the initial estimates may differ in different regions, this invention introduces a confidence weighting graph. This weight can be defined based on the initially estimated local gradient magnitude and the uncertainty of the feature map provided by the L-RSDformer (if available). A simplified definition is: in the initial estimated gradient, which is gentle (i.e., possibly a flat region) and corresponds to the original foggy image... Regions with simple textures are assigned a higher confidence level. The data fidelity term is defined in the form of a weighted L2 norm, as shown in formula (2):

[0030] Smoothing constraint This invention employs isotropic total variation (TV) regularization to suppress noise and force smooth changes in the transmission map across a continuous depth region. It effectively removes noise while maintaining piecewise smoothness, as shown in formula (3):

[0031] in, express In position Spatial gradient at a given location.

[0032] To avoid over-smoothing at object edges, this invention introduces an image content-guided edge-aware constraint. This utilizes the original foggy image. Edge information extracted (or intermediate features of L-RSDformer) is used to guide... The gradient direction. Specifically, the present invention aims to... The gradient can be larger at the edges of image textures and smaller in homogeneous regions. This invention employs a weighted anisotropic smoothing term, as shown in formula (4):

[0033] in, It is an edge indicator function, typically set with respect to the magnitude of the image gradient. Decreasing functions, for example ,parameter This makes it possible to [perform better] at strong edges in the image. The penalty for the gradient is reduced, thus allowing abrupt changes in the transmission map at this point.

[0034] For example, combining the above three factors, the complete problem of minimizing the energy of the refined transmission map is given by formula (5): This optimization problem is a convex optimization problem, which can be solved using efficient numerical algorithms, such as the primal-dual algorithm based on gradient descent or the alternating direction multiplier method (ADMM).

[0035] The optimal haze attenuation map is obtained by solving the problem. After that, you can pass A refined transmissive map is obtained. This transmissive map incorporates the depth and haze distribution priors provided by the L-RSDformer (through data terms) and satisfies the physical smoothness and edge preservation constraints, resulting in higher accuracy and spatial consistency compared to the initial estimate.

[0036] For example, the core idea of ​​the atmospheric light adaptive estimation algorithm is: to use the preliminary dehazing results of the L-RSDformer... and refined transmission map As a priori, the optimal atmospheric light value for each pixel location or local region is estimated by solving an optimization problem constrained by a physical model. Specific steps may include:

[0037] (1) Initialization: Based on the inverse process of the reformal atmospheric scattering model proposed in this invention, combined with the preliminary dehazing result J(x) output by L-RSDformer, the initial atmospheric light estimate is calculated. Then, combined with traditional prior corrections such as dark channel, a rough initial atmospheric light distribution map Ainit(x) is generated, providing a reasonable starting point for subsequent optimization.

[0038] (2) Constructing the objective function: Define the objective function based on the atmospheric light map The energy function, which includes a data fidelity term and a spatial smoothing constraint term, is shown in formula (6): (6) The first term (data fidelity term) mandates that the estimated atmospheric light must enable the reformulated scattering model to accurately reconstruct the observed foggy image. The second term (smoothing term) is based on the prior knowledge that atmospheric illumination varies continuously in space, suppressing drastic fluctuations in the estimation results. (Parameters...) Control the smoothing intensity.

[0039] (3) Solution and Optimization: The above optimization problem is a quadratic minimization problem. By discretizing it and constructing the corresponding system of linear equations, it can be solved using efficient numerical methods (such as the preprocessed conjugate gradient method). In each iteration, the algorithm fixes the transmission image. Update atmospheric light map This ensures that it evolves in a direction that satisfies the physical model and is spatially smooth.

[0040] (4) Physical boundary constraints: Boundary constraints are introduced during the solution process. ,in, The maximum possible pixel value of the image (e.g., 255) is used to ensure the physical plausibility of atmospheric light estimation.

[0041] For example, this invention can decouple the coupled optimization problem of haze attenuation map and atmospheric light map by using the ADMM framework, decomposing it into two main sub-problems: optimizing the haze attenuation map with a fixed atmospheric light map, and optimizing the atmospheric light map with a fixed haze attenuation map. Auxiliary variables are introduced to handle non-smooth terms, an augmented Lagrangian function is constructed, and iterative updates are performed according to the following steps:

[0042] (1) Update the haze attenuation map D: Solve the sparse linear system; (2) Update atmospheric light map A: Solve a constrained quadratic programming problem; (3) Update auxiliary variables u and v: use the soft threshold shrinkage operator; (4) Update the Lagrange multipliers μ and ν: use the standard gradient ascent step.

[0043] Through theoretical analysis and experimental verification, when the penalty parameters ρ1 and ρ2 are sufficiently large, the algorithm sequence converges to the global optimal solution, ensuring both optimization effect and efficiency.

[0044] For example, the present invention improves the original atmospheric scattering model by introducing a spatially varying atmospheric light map A(x) and a comprehensive haze influence factor ρ(x), where ρ(x) = β(x)・d(x) + δ(x); Where β(x) is the spatially varying scattering coefficient, d(x) is the scene depth, and δ(x) is the multiple scattering compensation term, the transmission graph t(x) is redefined as e (-ρ(x)) This yields a re-formulated atmospheric scattering model: I(x)=J(x)・e (-[β(x)d (x)+δ(x)]) +A(x)(1-e (-[β(x)d(x)+δ(x)]) ); Where I(x) is a haze image, J(x) is a clear image, and A(x) is an atmospheric light image.

[0045] The re-formulated atmospheric scattering model is rewritten in logarithmic domain form, and the haze attenuation map D(x) = -ln(t(x)) = ρ(x) is defined. A small constant ε is introduced to enhance numerical stability, and the estimated relationship between the normalized observation image and the haze attenuation map is obtained, which provides a basis for subsequent optimization.

[0046] For example, this invention integrates PMG-Defog as a post-processing module with L-RSDformer to form a collaborative paradigm of "data-driven coarse recovery + physical model fine correction," specifically including: First, L-RSDformer is used to perform preliminary dehazing on the haze image, and the coarse restoration result is output; Then, PMG-Defog takes the coarse restoration result and the original haze image as input and obtains a refined haze attenuation map through the above optimization process. With atmospheric light diagram A*(x); Finally, based on the optimal haze attenuation map, the refined transmittance map is determined using the following formula: T(x) = max(e (-D*(x)) , t0); Where t0 is the lower limit of the transmission map, and T(x) is the refined transmission map. This is the optimal haze attenuation map; The refined transmittance map and atmospheric light map are input into the reformulated atmospheric scattering model for inversion to obtain the final dehazed image, which may include: J refined (x)=(I(x)-A*(x)(1-T(x))) / T(x); Among them, J refined (x) is the final dehazed image, A*(x) is the optimized atmospheric light map, and T(x) is the refined transmission map.

[0047] Example 2 For example, when performing dehazing of remote sensing images, the original remote sensing image input to the L-RSDformer can be 512×512 pixels in size, with RGB three channels; the L-RSDformer outputs a coarse restoration result J(x): generated through a 4-level encoder-decoder structure, an MDTA attention module, and a GDFN feedforward network; its specific generation process is achieved collaboratively through a 4-level encoder-decoder U-shaped structure, an MDTA (Multi-Scale Dynamic Filter Attention) module, and a GDFN (Dynamically Gated Feedforward) network, and the specific data processing process of each module is as follows: (1) Input preprocessing module: The 512×512×3 original haze image I(x) normalized to the [0,1] interval is mapped to a 512×512×64 high-dimensional initial feature map F0 through two 3×3 convolutional layers with a stride of 1, completing the extraction of shallow texture and global structural features, laying the foundation for subsequent multi-scale coding.

[0048] (2) 4-level multi-scale coding module (including MDTA attention): The encoder adopts a 4-level downsampling structure, each level containing one MDTA attention module and one 2×2 downsampling convolution with a stride of 2: MDTA module: Generates spatially adaptive convolution kernels through dynamic filters to extract multi-scale frequency features; combined with lightweight window self-attention (8×8 window), it enhances the feature response of high-frequency ground feature details and low-frequency global haze distribution in remote sensing images with low parameter quantity; Downsampling: The resolution of the feature map is halved after each level of encoding (512×512→256×256→128×128→64×64→32×32), and the final output is 4-level multi-scale encoder features Fe1∼Fe4, where Fe1 is a high-resolution shallow detail feature and Fe4 is a low-resolution global haze distribution feature.

[0049] (3) DuRB: Global feature fusion is performed on the deepest feature Fe4 of the encoder. The channel attention mechanism is introduced to enhance the feature expression of global haze distribution and scene depth, and outputs 32×32×64 global purified feature Fb to provide global context constraints for the decoder.

[0050] (4) 4-level multi-scale decoding module (including GDFN feedforward network): The decoder adopts a 4-level upsampling structure, each level including bilinear interpolation upsampling, skip connections, and GDFN dynamic gated feedforward network: Skip connection: The corresponding scale features (Fe3~Fe1) of the encoder are concatenated with the upsampled feature channels of the decoder to compensate for the loss of details caused by downsampling; GDFN module: Generates spatial adaptive fusion weights through dynamic gating mechanism, balances shallow details of encoder and deep semantic features of decoder, suppresses redundant information, and enhances the transmission of effective dehazing features; Upsampling process: After step-by-step upsampling, the feature map resolution is restored step by step (32×32→64×64→128×128→256×256→512×512), and the final output is a 512×512×64 enhanced feature map Fout.

[0051] (5) Output reconstruction module: Fout is mapped to a 512×512×3 feature map through two 3×3 convolutional layers. The residual connection is introduced to add the original haze image I(x) to the output residual, and finally output the coarse recovery result J(x), completing the initial dehazing of L-RSDformer.

[0052] Key parameter settings for PMG-Defog may include ε=10−3, λs=0.05, λe=0.1, η=0.01, γ=0.03, ρ1=10, ρ2=10, t0=0.1; Number of iterations: 35; For example, the execution flow of the present invention in performing remote sensing image dehazing may include: (1) Input preprocessing: Normalize the original haze remote sensing image I(x) to the [0,1] interval and input it into the L-RSDformer model; (2) Coarse recovery stage: L-RSDformer extracts multi-scale frequency features through dynamic filters, and outputs coarse recovery image J(x) after dynamic gating fusion; (3) Initial parameter estimation: Based on J(x) and the original image I(x), calculate the initial atmospheric light Ainit(x) (combined with the dark channel prior) and the initial haze attenuation map Dinit(x); (4) Multi-stage optimization: (a) Fix Ainit(x) and solve the transmission map refinement energy function to obtain D(1)(x); (b) Fix D(1)(x) and solve the atmospheric light adaptive estimation energy function to obtain A(1)(x); Repeat the iteration 35 times, updating D(x), A(x), auxiliary variables, and Lagrange multipliers, to finally obtain... With A*(x); (5) Final defogging: Substitute into the inversion formula to calculate J refined (x), outputting the final fog-free image.

[0053] For example, on the RICE-I dataset (500 pairs of remote sensing images), the PSNR of the final dehazed image obtained by cascading the PMG-Defog model and L-RSDformer reached 29.42 dB, which is 0.77 dB higher than that obtained by using L-RSDformer alone; the FSIM reached 0.980, which is 0.7 percentage points higher; and the processing time for a single 512×512 image was about 0.8 seconds, which meets the practical efficiency requirements.

[0054] The above are remote sensing image dehazing methods provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding remote sensing image dehazing system, including: The acquisition module is used to acquire the original haze image of the area to be tested; The preliminary dehazing module is used to perform preliminary dehazing on the original haze image using a pre-trained lightweight deep learning model, L-RSDformer, to determine the preliminary dehazing result. The lightweight deep learning model L-RSDformer includes an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer connected in sequence. Specifically, the input layer maps the original haze image to high-dimensional initial features; the multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features; the dual residual bottleneck layer (DuRB) performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features; the multi-scale decoding layer upsamples the global features to determine enhanced features; and the output layer maps the enhanced features to the preliminary dehazing result. The dehazing module takes the preliminary dehazing results and the original haze image as input to the constructed multi-stage optimization mathematical model PMG-Defog. Based on the preliminary dehazing results and the original haze image, it determines the initial haze attenuation map and the initial atmospheric light map. By fixing the initial atmospheric light map, it minimizes the refined energy function of the transmission map to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, it minimizes the adaptive estimation energy function of atmospheric light to determine the atmospheric light value for the current iteration. The iteration process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing a comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model (ASM) is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazed image.

[0055] Specific limitations regarding remote sensing image dehazing systems can be found in the limitations of remote sensing image dehazing methods described above, and will not be repeated here. Each module in the aforementioned remote sensing image dehazing system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0056] The present invention also provides a computer-readable storage medium storing a computer program that can be used to perform the above-described remote sensing image dehazing.

[0057] The present invention also provides Figure 4 The schematic diagram of the computer device shown is as follows: Figure 4As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the remote sensing image dehazing method provided in the above embodiments.

[0058] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0059] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for dehazing remote sensing images, characterized in that, include: Acquire the original haze image of the area to be tested; A pre-trained lightweight deep learning model, L-RSDformer, is used to perform preliminary dehazing on the original haze image to determine the preliminary dehazing result. The L-RSDformer model comprises an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer, connected sequentially. The input layer maps the original haze image to high-dimensional initial features. The multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features. The DuRB performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features. The multi-scale decoding layer upsamples the global features to determine enhanced features. The output layer maps the enhanced features to the preliminary dehazing result. The preliminary dehazing results and the original haze image are used as inputs to the constructed multi-stage optimization mathematical model PMG-Defog. Based on the preliminary dehazing results and the original haze image, the initial haze attenuation map and the initial atmospheric light map are determined. By fixing the initial atmospheric light map, the energy function of the transmission map is minimized to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, the adaptive estimation energy function of atmospheric light is minimized to determine the atmospheric light value for the current iteration. The iterative process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing a comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model (ASM) is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazing image.

2. The remote sensing image dehazing method as described in claim 1, characterized in that, The acquisition of the re-formulated atmospheric scattering model specifically includes: The comprehensive haze impact factor ρ(x) is obtained using the following formula: ρ(x) = β(x)・d(x) + δ(x); Where β(x) is the spatially varying scattering coefficient, d(x) is the scene depth, and δ(x) is the multiple scattering compensation term. Based on the comprehensive haze impact factor ρ(x), and combined with the spatially varying atmospheric light map A(x), the re-formulated atmospheric scattering model is obtained using the following formula: I(x)=J(x)・e (-[β(x)d (x)+δ(x)]) +A(x)(1-e (-[β(x)d(x)+δ(x)]) ); Where I(x) is the haze image, J(x) is the clear image, A(x) is the atmospheric light map, and the transmission map t(x) = e (-ρ(x)) .

3. The remote sensing image dehazing method as described in claim 1, characterized in that, The process of determining a refined transmittance map based on the optimal haze attenuation map specifically includes: Based on the optimal haze attenuation map, the refined transmittance map is determined using the following formula: T(x)=max(e (-D*(x)) ,t0); Where t0 is the lower limit of the transmission map, T(x) is the refined transmission map, and D * (x) is the optimal haze attenuation map.

4. The remote sensing image dehazing method as described in claim 3, characterized in that, The process of inputting the refined transmission map and the optimized atmospheric light map into the re-formulated atmospheric scattering model for inversion to obtain the final dehazed image specifically includes: J refined (x)=(I(x)-A*(x)(1-T(x))) / T(x); Among them, J refined (x) is the final dehazed image, A*(x) is the optimized atmospheric light map, and T(x) is the refined transmission map.

5. The remote sensing image dehazing method as described in claim 1, characterized in that, The refined energy function of the transmission map is determined by introducing regularization parameters to weight the obtained smoothing constraint term and edge preservation term, summing the obtained data fidelity term, the weighted smoothing constraint term and edge preservation term.

6. The remote sensing image dehazing method as described in claim 1, characterized in that, The acquisition of the atmospheric light adaptive estimation energy function specifically includes: The atmospheric light adaptive estimation energy function is obtained as follows: Where I(x) is the haze image and A(x) is the atmospheric light image. To smooth the intensity control parameters, For preliminary defogging results, For a refined transmission image.

7. The remote sensing image dehazing method as described in claim 1, characterized in that, Before performing preliminary defogging on the original haze image, the original haze image is normalized.

8. A remote sensing image dehazing system, characterized in that, include: The acquisition module is used to acquire the original haze image of the area to be tested; The preliminary dehazing module is used to perform preliminary dehazing on the original haze image using a pre-trained lightweight deep learning model, L-RSDformer, to determine the preliminary dehazing result. The lightweight deep learning model L-RSDformer includes an input layer, a multi-scale encoding layer, a dual residual bottleneck layer (DuRB), a multi-scale decoding layer, and an output layer connected in sequence. Specifically, the input layer maps the original haze image to high-dimensional initial features; the multi-scale encoding layer extracts multi-scale encoder features from the high-dimensional initial features; the dual residual bottleneck layer (DuRB) performs global fusion and channel attention enhancement on the multi-scale encoder features to determine global features; the multi-scale decoding layer upsamples the global features to determine enhanced features; and the output layer maps the enhanced features to the preliminary dehazing result. The dehazing module takes the preliminary dehazing results and the original haze image as input to the constructed multi-stage optimization mathematical model PMG-Defog. Based on the preliminary dehazing results and the original haze image, it determines the initial haze attenuation map and the initial atmospheric light map. By fixing the initial atmospheric light map, it minimizes the refined energy function of the transmission map to determine the haze attenuation map for the current iteration. By fixing the haze attenuation map for the current iteration, it minimizes the adaptive estimation energy function of atmospheric light to determine the atmospheric light value for the current iteration. The iteration process is repeated until convergence to obtain the optimal haze attenuation map and the optimized atmospheric light map. Based on the optimal haze attenuation map, the refined transmission map is determined. By introducing a comprehensive haze influence factor and the atmospheric light map, the atmospheric scattering model (ASM) is improved to determine the re-formalized atmospheric scattering model. The refined transmission map and the optimized atmospheric light map are input into the re-formalized atmospheric scattering model for inversion to obtain the final dehazed image.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the remote sensing image dehazing method according to any one of claims 1-7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the remote sensing image dehazing method according to any one of claims 1-7.

Images