Mamb-Conv mixed multi-scale fusion remote sensing image color correction method and system

Abstract

In order to solve the problem of color cast of a remote sensing image caused by various factors, the invention provides a Mamb-Conv mixed multi-scale fusion remote sensing image color correction method and system. The invention provides a Mama-Conv hybrid multi-scale fusion network (MaCoFusion), the network is combined with a Mama-Conv (MaCo) module and a multi-scale adaptive feature fusion (MAFF) module, and through the efficient sequence modeling capability of a Mama model and the local feature extraction capability of convolution operation, the global and local dependency relationship of an image is captured. And the MAFF module adaptively fuses multi-scale features through a channel grouping scale attention mechanism (CGSA), so that the color fidelity and the visual authenticity are improved. The method supports the input image of any size, and has excellent effects in the aspects of color restoration, structure maintenance and the like compared with the existing method, and the interpretability and the application value of the remote sensing image are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing, and particularly relates to a color correction model based on deep learning for color cast of remote sensing images. BACKGROUND

[0002] Remote Sensing Images (RSIs) are acquired by satellites, aerial platforms or unmanned aerial vehicles, and are widely used in environmental monitoring, urban planning, agricultural management, disaster assessment and military reconnaissance. However, due to various external factors in the imaging process, such as atmospheric scattering, illumination changes and device calibration errors, remote sensing images often exhibit color cast, which not only reduces the visual quality of the images, but also significantly affects the performance of subsequent visual tasks, such as semantic segmentation, object detection and land use classification. The goal of remote sensing image color correction is to restore the true colors of the images through algorithms or models, thereby improving the image quality and application value.

[0003] Traditional remote sensing image color correction methods mainly fall into two categories: physical models and statistical methods. Physical models are based on radiative transfer theory, which simulates the propagation of light in the atmosphere to correct scattering and absorption effects. Typical models include 6S, ATCOR and MODTRAN. These methods have strong theoretical interpretability, but rely on accurate atmospheric parameters (such as aerosol optical depth, atmospheric water vapor content), which are difficult to obtain in real time, dynamic and costly. Statistical methods include histogram matching, Wallis transformation or optimal transport mapping to achieve color transfer. These methods do not require complex atmospheric modeling, but have poor adaptability in cloud coverage, complex topography or large-scale image scenes, and often require high-quality reference images as support.

[0004] With the rise of deep learning, Convolutional Neural Networks (CNNs) have been widely applied in image color correction. CNNs significantly improve the correction accuracy through their powerful feature extraction capabilities. For example, CNN models estimate global illumination values and then perform color correction based on the Retinex theory. End-to-end models based on the U-Net architecture perform well in underwater image correction. However, CNNs have limitations in capturing long-range dependencies and multi-scale features, which can lead to over-saturation, under-correction or loss of details in complex remote sensing scenes. In recent years, the Transformer architecture has achieved global modeling through self-attention mechanisms, but its quadratic computational complexity limits its practicality when dealing with large-scale remote sensing images. SUMMARY

[0005] To address the shortcomings of existing technologies, this paper proposes a Mamba-Conv hybrid multi-scale fusion network (MaCoFusion). Combining the advantages of state-space models and convolution operations, this invention improves image fidelity and perceptual quality through adaptive multi-scale feature fusion, effectively correcting color casts in remote sensing images while maintaining computational efficiency and adaptability to diverse scenarios.

[0006] The technical solution adopted by the present invention is: a Mamba-Conv hybrid multi-scale fusion remote sensing image color correction method, which specifically includes the following steps: Based on the existing remote sensing image dataset, normal images are screened and color-biased images are synthesized using color constancy theory to form a large-scale paired dataset. Construct a color correction network, namely the MaCoFusion network, which includes the MaCo module and the multi-scale adaptive feature fusion MAFF module. The MaCo module combines the long sequence modeling capability of Mamba and the local feature extraction capability of convolution to capture global and local features; the MAFF module includes spatial attention and channel grouping scale attention. The specific processing process is as follows: First, the input image is converted into a multi-scale image pyramid. At each scale, the image is converted into a feature representation and then input into multiple MaCo modules for processing. The features output by the MaCo modules are aligned with features of other scales through upsampling or downsampling, and feature splicing is performed through cross-scale connections. Then, the multi-scale fusion features are generated by the MAFF module, and the multi-scale fusion features are divided into several correction components along the channel dimension. Each component is successively added to the full-scale input image through residual connections to generate a corrected image. The color-cast remote sensing image is input into the trained MaCoFusion network and the color-corrected image is output.

[0007] Furthermore, the paired dataset is formed as follows: (1) Normal image screening: crop the images in the dataset and convert them from RGB color space to CIELAB color space to extract brightness. and chrominance components 、 , the color shift factor is calculated using the equivalent circle model:

[0008]

[0009]

[0010] in for Channel mean, for Channel chromaticity center distance, and They are The channel mean and chromaticity center distance are calculated in the same way; is the color shift distance, is the equivalent circle radius, is the color cast distortion factor, The larger the value, the more obvious the color shift; the image meets considered as a candidate for normal imaging; Then, a secondary verification is performed using the grayscale index of the no-reference color correction algorithm. If the color difference between the corrected image and the original image is less than 2, the image is confirmed to have no significant color difference. After that, images with clouds or black edges are manually screened out to ensure the quality of the dataset. (2) Color-biased image synthesis: According to the color constancy theory, the color-biased image is modeled as the product of the normal image and the color-biased illumination:

[0011] in It is a color cast image. is the color cast lighting matrix, It's a normal image. The range is obtained based on the color characteristics of real remote sensing images; (3) The normal / biased color paired data are used as paired datasets.

[0012] Furthermore, the input image is downsampled by bilinear interpolation to generate a four-scale image pyramid, namely full scale, 1 / 2 scale, 1 / 4 scale and 1 / 8 scale, which are expressed as: (1) in, is the input image, For the A scale image.

[0013] Furthermore, at each scale, the image is converted into feature representation through the convolution layer. The MaCo module first stabilizes the feature distribution through layer normalization LayerNorm, and the normalized features are reshaped into a sequence form. , input Mamba block, where H, W, C are the height, width and number of channels of the feature map respectively; in Mamba block, each row in the sequence is regarded as a time step The input, that is , whose calculation is defined as: (2) (3) in, is the time step The eigenvector of is hidden state, For output, 、 、 is the state transition matrix; then, the sequence characteristics Reshape into a two-dimensional feature map through convolution and ReLU activation operations ; To achieve cross-scale information interaction, the features output by the MaCo module are aligned with features of other scales through upsampling or downsampling, and feature splicing is performed through cross-scale connections; MaCo blocks at scale The input-output relationship is expressed as: (4) in, represents feature splicing, and represent downsampling and upsampling respectively, Indicates that downsampling or upsampling is selected based on the position of the module in the network.

[0014] Furthermore, the MAFF module receives features from two different scales: larger scale features and smaller scale features ,The MaCoFusion network contains three MAFF modules, each module receives two scales, one of which receives the full-scale and 1 / 4-scale image features fused into features , and the other receives 1 / 2 scale and 1 / 8 scale image features fused as features , the last one to receive and Further integration.

[0015] Furthermore, the spatial attention processing process in the MAFF module is as follows: Larger scale features and smaller scale features Perform maximum pooling and average pooling on the features along the channel dimension respectively, concatenate the two pooling results and generate a single channel attention map through the convolution layer. After activation, it is multiplied with the input feature map to generate an enhanced feature map and ; The processing of spatial attention is expressed as: (5) (6) in, for function, denotes convolution, denotes feature concatenation, denotes element-wise multiplication, denotes input feature, denotes max-pooling, denotes average-pooling; subsequently, down-sampling by two convolutional layers to generate where , is 1 / 8 of the original size; generate the same size by the same operation; introduce the scale dimension, concatenate and along the scale dimension to form the scale concatenation feature .

[0016] Further, in the channel grouping scale attention, is divided into groups, each group is denoted as , and the attention weight of each group is calculated as: (7) where is the representation of in the th group, and is a scaling factor to prevent numerical instability; the weight is expanded to channels, and after the group weights are concatenated along the channel dimension, the attention weight is normalized along the scale dimension by the function to generate the attention weight , and the fused feature is calculated as: (8) subsequently, up-sampling to the original resolution by bilinear interpolation, and adding it to by residual connection to generate the multi-scale fused feature : .

[0017] Further, the multi-scale fused feature is divided into corrective components each component is added to the full-scale input image in sequence through a residual connection to generate a corrected image: (10) wherein, is the full-scale input image, is the corrected image.

[0018] Further, a composite loss function is used in the training process, combining loss and perceptual loss: (11) wherein, is the pixel-level difference loss, used to ensure correction accuracy; is the perceptual loss based on the pre-trained VGG16 network, used to enhance perceptual quality, and the features of ReLU2_2, ReLU3_3 and ReLU4_3 are used for calculation, and the calculation method is to input the reconstructed image and the target image into VGG16 respectively, extract the feature maps of the corresponding layers, and then calculate the distance between them;

[0019] wherein denotes the target image, denotes the reconstructed image, denotes the feature map extracted by VGG16 in the l layer, denotes the weight of each layer.

[0020] The application also provides a Mamba-Conv hybrid multi-scale fusion remote sensing image color correction system, comprising: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the storage instructions in the memory to execute the high-resolution remote sensing image change detection method based on decoupling representation learning as described in the above technical solution.

[0021] Compared with existing technologies, the present invention offers the following advantages and benefits: It proposes a remote sensing image color correction method based on the Mamba-Conv hybrid multi-scale fusion network (MaCoFusion). This network combines the MaCo module and the MAFF module, leveraging Mamba's efficient sequence modeling capabilities with the local feature extraction capabilities of convolution operations to achieve global and local feature fusion. The MAFF module adaptively fuses multi-scale features through spatial attention and channel-grouped scale attention, improving color fidelity and visual realism. This method accepts input images of any size and, compared to existing methods, outperforms both quantitative color correction metrics (such as PSNR, CD, and MS-SWD) and visual quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the MaCoFusion network model; Figure 2 It is a MaCo module structure; Figure 3 It is a MAFF module structure; Figure 4 Color correction results for three scenes in the GID-5 dataset; Figure 5 Color correction results for three scenes in the LoveDA dataset; Figure 6 Color correction results of two scenes in real color cast images; DETAILED DESCRIPTION The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0023] The present invention belongs to remote sensing image processing technology, and in particular to a color correction model based on deep learning for the color cast problem of remote sensing images. The network model structure diagram is as follows Figure 1 、 Figure 2 and Figure 3 The present invention can use a computer to train and infer the network model, and use the PyTorch deep learning framework under the Linux operating system. The specific experimental environment configuration is shown in Table 1.

[0024] Table 1 Experimental environment configuration

[0025] The embodiment of the present invention provides a Mamba-Conv hybrid multi-scale fusion remote sensing image color correction method, which is specifically implemented as follows: Step 1: Dataset Construction. The dataset used in the experiment is based on the GID-5 and LoveDA datasets. The GID-5 dataset contains 150 high-quality images (4-meter spatial resolution) covering diverse geographic regions and land cover types across China. The LoveDA dataset contains 5,987 high-resolution (0.3-meter spatial resolution) images of urban and rural areas. Before training, the image data was processed as follows to obtain paired data with normal / discolored images.

[0026] (1) Normal image screening. The images in the GID-5 dataset were cropped to 512×512 pixels, and the LoveDA dataset was kept at the original size of 1024×1024 pixels. The images were converted from RGB color space to CIELAB color space and the brightness was extracted. and chrominance components 、 The color shift factor is calculated using the equivalent circle model:

[0027]

[0028]

[0029] in for Channel mean, for Channel chromaticity center distance, and They are The channel mean and chromaticity center distance are calculated in the same way as above. is the color shift distance, is the equivalent circle radius, is the color cast distortion factor, The larger the value, the more obvious the color shift. These images were considered candidates for normal status. To improve screening reliability, a secondary verification was performed using the grayness index (GI) of a reference-free color correction algorithm. If the color difference between the corrected image and the original image was less than 2, the image was considered to have no significant color difference. Images containing clouds or black edges were then manually screened out to ensure dataset quality.

[0030] (2) Color-biased image synthesis. According to the color constancy theory, a color-biased image can be modeled as the product of a normal image and color-biased illumination:

[0031] in It is a color cast image. is the color cast lighting matrix, It is a normal image. The range is obtained according to the color characteristics of the real remote sensing image to ensure the authenticity of the synthesized data.

[0032] After synthesis, 2248 normal / abnormal image pairs are obtained for the GID-5 dataset, and 3561 image pairs are obtained for the LoveDA dataset.

[0033] Step 2: Network construction: design the MaCoFusion network, which includes the Mamba-Conv (MaCo) module and the multi-scale adaptive feature fusion (MAFF) module. The MaCo module combines the long sequence modeling capability of Mamba and the local feature extraction capability of convolution to capture global and local features; the MAFF module adaptively fuses multi-scale features through the channel grouped channel attention (CGSA) mechanism. The MaCoFusion network takes the abnormal color remote sensing image as input, and outputs the color corrected image through multi-scale feature extraction, interaction and fusion. The specific construction steps include: Step 21: Build an image pyramid. The input image is down-sampled by bilinear interpolation to generate a four-scale image pyramid, which is full scale, 1 / 2 scale, 1 / 4 scale and 1 / 8 scale, represented as: (1) wherein, is the input image, is the image of the th scale.

[0034] Step 22: Multi-scale feature extraction and interaction. At each scale, the image is converted to a feature representation by a convolution layer, where H, W and C are the height, width and number of channels of the feature map, respectively, and then input into the MaCo module for processing. The MaCo module first stabilizes the feature distribution through layer normalization (LayerNorm) to alleviate the gradient problem. The normalized feature is reshaped into a sequence form and input into the Mamba block. In the Mamba block, each row in the sequence is treated as the input of time step , i.e. , whose calculation is defined as: (2) (3) wherein, is the feature vector of time step , is the hidden state, is the output, 、 、 is the state transition matrix. Mamba’s dynamic selection mechanism is based on the input Adjustment 、 、 , effectively capturing long-range dependencies and enhancing global context modeling. Subsequently, sequence features Reshape into a two-dimensional feature map through convolution and ReLU activation operations .

[0035] To achieve cross-scale information interaction, the features output by the MaCo module are aligned with features of other scales through upsampling or downsampling, and feature concatenation is performed through cross-scale concatenation (CSC). MaCo blocks at scale The input-output relationship is expressed as: (4) in, represents feature splicing, and represent downsampling and upsampling respectively, Indicates the selection of downsampling or upsampling based on the module's position in the network. By constructing a multi-scale image pyramid and cross-scale feature splicing, the network can simultaneously process high-resolution details and low-resolution global context, enhancing feature expression capabilities.

[0036] Step 23: Adaptive multi-scale feature fusion. Multi-scale adaptive feature fusion is led by the MAFF module, which includes spatial attention (SA) and channel group scale attention (CGSA). Figure 1 As shown in , the MAFF module receives features from two different scales: the larger scale features and smaller scale features , the network contains three MAFF modules, each module receives two scales, one of which is a larger scale and the other is a smaller scale, that is, there are three MAFF modules in the network, one of which receives the full scale and 1 / 4 scale fusion as features , the other receives 1 / 2 scale and 1 / 8 scale fusion as features , the last one to receive and First, the feature representation is enhanced by spatial attention, specifically by performing maximum pooling and average pooling on the features along the channel dimension, and then the two pooling results are concatenated and passed through the convolution layer to generate a single channel attention map. After activation, it is multiplied with the input feature map to generate an enhanced feature map and , the process can be expressed as: (5) (6) in, for function, express convolution, Represents element-wise multiplication.

[0037] Then, Generated by downsampling through two convolutional layers ,in , 1 / 8 of the original size. Generate the same size through similar operations . Introducing the scale dimension, and Along the scale dimension Splicing to form scale splicing features ,in .

[0038] In CGSA, Divided into groups, each group is represented by , The attention weight of each group is calculated as: (7) in, for In the The representation in the group, To prevent numerical instability, the scaling factor is set. Expand to aisle, After the group weights are concatenated along the channel dimension, The function is normalized along the scale dimension to generate attention weights The fusion feature is calculated as: (8) Then, Upsampled to native resolution via bilinear interpolation , connected with Add to generate multi-scale fusion features : (9) Step 24: Correction component generation. Split along the channel dimension into Correction Component , each component has the same number of channels as the input image, and is added to the full-scale input image through residual connections to generate a corrected image: (10) Step 25: Construct loss function. To optimize the color correction performance, a composite loss function is used, combining Loss and Perceived Loss: (11) in, Pixel-level differences Loss, used to ensure correction accuracy; To enhance the perceptual quality based on the perceptual loss of the pre-trained VGG16 network, the present invention uses the features of the three layers ReLU2_2, ReLU3_3 and ReLU4_3 for calculation. The calculation method is to input the reconstructed image and the target image into VGG16 respectively, extract the feature maps of the corresponding layers, and then calculate the relationship between them. distance;

[0039] in represents the target image, represents the reconstructed image, Indicates that VGG16 is in the l The feature map extracted by the layer, Represents the weight of each layer.

[0040] Step 3: Model training. During model training, randomly cropped 256×256 pixel patches from the color-distorted image as input, and cropped patches at the same location from the normal image as the ground truth (GT). The training and test sets were split in a 7:3 ratio. Training lasted 100 epochs with a batch size of 8. The AdamW optimizer was used, and the initial learning rate was . , the weight decay is .

[0041] Step 4: Model inference: During model inference, a remote sensing image of any size is input, and after being processed by the model, a rectified image of the same size is output.

[0042] The images corrected using this method and the qualitative comparison with other methods are shown in Figure 2. Figure 4 、 Figure 5 and Figure 6 shown. Figure 4 The results on the GID-5 dataset are: Figure 5 The results on the LoveDA dataset are: Figure 6The results of the color cast image without corresponding ground truth (GT). It can be seen that the method can accurately correct color distortion while maintaining the appearance of natural images, and is very suitable for color correction in complex remote sensing scenes. The quantitative evaluation results are shown in Table 2. On the GID-5 and LoveDA data sets, MaCoFusion achieved the best performance in the peak signal-to-noise ratio (PSNR) and color difference (CD) indicators.

[0043] Table 2. Quantitative comparison results of different methods on GID-5 and LOVEDA data sets.

[0044]

[0045] In another aspect, the embodiments of the present application also provide a Mamba-Conv hybrid multi-scale fusion remote sensing image color correction system, comprising: The processor and the memory, the memory is used for storing program instructions, and the processor is used for calling the storage instructions in the memory to execute the high-resolution remote sensing image change detection method based on decoupling representation learning as described in the above technical solutions.

[0046] The specific embodiments described herein are merely illustrative of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

Claims

1. A Mamba-Conv hybrid multi-scale fusion remote sensing image color correction method, characterized in that: The steps include: Based on the existing remote sensing image dataset, normal images are screened and color-biased images are synthesized using color constancy theory to form a large-scale paired dataset. Construct a color correction network, namely the MaCoFusion network, which includes the MaCo module and the multi-scale adaptive feature fusion MAFF module. The MaCo module combines the long sequence modeling capability of Mamba and the local feature extraction capability of convolution to capture global and local features; the MAFF module includes spatial attention and channel grouping scale attention. The specific processing process is as follows: First, the input image is converted into a multi-scale image pyramid. At each scale, the image is converted into a feature representation and then input into multiple MaCo modules for processing. The features output by the MaCo modules are aligned with features of other scales through upsampling or downsampling, and feature splicing is performed through cross-scale connections. Then, the multi-scale fusion features are generated by the MAFF module, and the multi-scale fusion features are divided into several correction components along the channel dimension. Each component is successively added to the full-scale input image through residual connections to generate a corrected image. The color-biased remote sensing image is input into the trained MaCoFusion network and the color-corrected image is output.

2. The remote sensing image color correction method using Mamba-Conv hybrid multi-scale fusion as claimed in claim 1, characterized in that: The implementation of forming a paired dataset is as follows: (1) Normal image screening: crop the images in the dataset and convert them from RGB color space to CIELAB color space to extract brightness. and chrominance components 、 , the color shift factor is calculated using the equivalent circle model: in for Channel mean, for Channel chromaticity center distance, and They are The channel mean and chromaticity center distance are calculated in the same way; is the color shift distance, is the equivalent circle radius, is the color cast distortion factor, The larger the value, the more obvious the color shift; the image meets considered as a candidate for normal imaging; Then, a secondary verification is performed using the grayscale index of the no-reference color correction algorithm. If the color difference between the corrected image and the original image is less than 2, the image is confirmed to have no significant color difference. After that, images with clouds or black edges are manually screened out to ensure the quality of the dataset. (2) Color-biased image synthesis: According to the color constancy theory, the color-biased image is modeled as the product of the normal image and the color-biased illumination: in It is a color cast image. is the color cast lighting matrix, It's a normal image. The range is obtained based on the color characteristics of real remote sensing images; (3) The normal / biased paired data are used as paired datasets.

3. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 1, characterized in that: The input image is downsampled by bilinear interpolation to generate a four-scale image pyramid, namely full scale, 1 / 2 scale, 1 / 4 scale and 1 / 8 scale, which are expressed as: （1） in, is the input image, For the A scale image.

4. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 1, characterized in that: At each scale, the image is converted into feature representation through the convolution layer. The MaCo module first stabilizes the feature distribution through layer normalization LayerNorm. The normalized features are reshaped into a sequence form. , input Mamba block, where H, W, C are the height, width and number of channels of the feature map respectively; in Mamba block, each row in the sequence is regarded as a time step The input, that is , which is calculated as: （2） （3） in, is the time step The eigenvector of is hidden state, For output, 、 、 is the state transition matrix; then, the sequence characteristics Reshape into a two-dimensional feature map through convolution and ReLU activation operations ; To achieve cross-scale information interaction, the features output by the MaCo module are aligned with features of other scales through upsampling or downsampling, and feature splicing is performed through cross-scale connections; MaCo blocks at scale The input-output relationship is expressed as: （4） in, represents feature splicing, and represent downsampling and upsampling respectively, Indicates that downsampling or upsampling is selected based on the position of the module in the network.

5. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 1, characterized in that: The MAFF module receives features from two different scales: larger scale features and smaller scale features ,The MaCoFusion network contains three MAFF modules, each module receives two scales, one of which receives the full-scale and 1 / 4-scale image features fused into features , and the other receives 1 / 2 scale and 1 / 8 scale image features fused as features , the last one to receive and Further integration.

6. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 5, characterized in that: The processing of spatial attention in the MAFF module is as follows: Larger scale features and smaller scale features Perform maximum pooling and average pooling on the features along the channel dimension respectively, concatenate the two pooling results and generate a single channel attention map through the convolution layer. After activation, it is multiplied with the input feature map to generate an enhanced feature map and ; The processing of spatial attention is expressed as: （5） （6） in, for function, express convolution, represents feature splicing, represents element-wise multiplication, represents the input features, represents maximum pooling, represents average pooling; Then, Generated by downsampling through two convolutional layers ,in , 1 / 8 of the original size; Generate the same size by the same operation ; Introduce the scale dimension, and Along the scale dimension Splicing to form scale splicing features .

7. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 6, characterized in that: In channel grouping scale attention, Divided into groups, each group is represented by , , the attention weight of each group is calculated as: （7） in, for In the The representation in the group, To prevent numerical instability of the scaling factor; the weight Expand to aisle, After the group weights are concatenated along the channel dimension, The function is normalized along the scale dimension to generate attention weights , the fusion feature is calculated as: （8） Then, Upsampled to native resolution via bilinear interpolation , connected with Add to generate multi-scale fusion features : 。 8. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 1, characterized in that: Multi-scale fusion features Split along the channel dimension into Correction Component , each component has the same number of channels as the input image, and is added to the full-scale input image through residual connections to generate a corrected image: （10） in, is the full-scale input image, To correct the image.

9. The remote sensing image color correction method of Mamba-Conv hybrid multi-scale fusion according to claim 1, characterized in that: The composite loss function is used in the training process, combined with Loss and Perceived Loss: （11） in, Pixel-level differences Loss, used to ensure correction accuracy; It is a perceptual loss based on the pre-trained VGG16 network, used to enhance the perceptual quality. The features of the three layers ReLU2_2, ReLU3_3 and ReLU4_3 are used for calculation. The calculation method is to input the reconstructed image and the target image into VGG16 respectively, extract the feature maps of the corresponding layers, and then calculate the relationship between them. distance; in represents the target image, represents the reconstructed image, Indicates that VGG16 is l The feature map extracted by the layer, Represents the weight of each layer.

10. A Mamba-Conv hybrid multi-scale fusion remote sensing image color correction system, characterized in that: include: A processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a Mamba-Conv hybrid multi-scale fusion remote sensing image color correction method as described in any one of claims 1 to 9.

Images