Dark-light-oriented color-monochromatic double-camera unmanned aerial vehicle aerial photography imaging method

Abstract

The invention relates to a dark-light-oriented color-monochromatic dual-camera unmanned aerial vehicle aerial photography imaging method, and belongs to the field of low-light aerial photography imaging. The method comprises the following steps: acquiring paired low-illumination data sets at low altitude at night by using a color camera and a monochromatic camera carried by aerial photography equipment; constructing an integral structure of the aerial low-light imaging method ALICE based on the color-monochrome camera; constructing a disturbance corrector based on wavelet transform, performing rough correction on a degradation region of a low-light color image in the paired low-illumination data set to obtain a corrected image, and meanwhile, avoiding loss of specific texture information in an aerial scene; and constructing a refinement restorer based on monochromatic image guidance in ALICE, performing refinement restoration on the obtained corrected image, and performing fine restoration of exposure and details by utilizing the low-light monochromatic image guidance in the paired low-illumination data set to generate a high-quality aerial image with good exposure and rich texture details. According to the invention, a high-quality aerial image with good exposure and rich texture details can be accurately generated.

Description

Technical Field

[0001] The present invention belongs to the field of low-light aerial imaging, and particularly relates to a color-monochrome dual-camera unmanned aerial vehicle (UAV) aerial imaging method for low-light conditions. Background Art

[0002] Aerial imaging technology plays an important role in obtaining clear information of ground scenes. Compared with ground images, aerial images can capture richer spatial and color information and are widely used in many fields such as urban planning, UAV target tracking, agricultural monitoring, and small target detection. However, in low-light environments at night, due to the limited number of photons received by the camera sensor, low-light aerial images are prone to degradation phenomena such as under-exposure, noise enhancement, increased artifacts, and color distortion, seriously affecting the image quality.

[0003] The degradation of low-light aerial images not only reduces the visibility of the images but also poses significant challenges to subsequent computer vision tasks, resulting in a decrease in the accuracy of high-level vision tasks such as target recognition and object tracking. Therefore, enhancing, denoising, and color correction of low-light aerial images have become key research directions for improving the quality of night aerial imaging. An effective low-light aerial imaging method can improve the overall brightness of the image, suppress noise, and restore real color information, thereby enhancing the practicality of aerial images and providing more reliable visual data support for automated detection, monitoring, and analysis. Summary of the Invention

[0004] The purpose of the present invention is to make up for the gaps and deficiencies in the prior art and provide a color-monochrome dual-camera UAV aerial imaging method for low-light conditions, which can generate aerial images with good exposure and rich texture details.

[0005] To achieve the above purpose, the technical solution of the present invention is: a color-monochrome dual-camera UAV aerial imaging method for low-light conditions, comprising the following steps:

[0006] Step S1: Use the color camera and monochrome camera carried by the aerial imaging device to obtain the required paired low-illumination datasets at low altitude at night;

[0007] Step S2: Construct the overall structure of the aerial low-light imaging method ALICE based on the color-monochrome cameras;

[0008] Step S3: Construct a perturbation corrector based on wavelet transform in ALICE to roughly correct the degraded regions of the low-light color images in the paired low-illumination datasets obtained in Step S1 to obtain corrected images, while avoiding the loss of unique texture information in the aerial scene;

[0009] Step S4: Build a refinement restorer based on monochromatic image guidance in ALICE, refine and restore the corrected image obtained in step S3, and use the low-light monochromatic images in the paired low-light dataset to guide the fine restoration of exposure and details, generating high-quality aerial images with good exposure and rich texture details.

[0010] Further, step S1 is specifically as follows:

[0011] Step S11: In the low-altitude night environment, use the color camera carried by the aerial photography device to collect low-light color images with short exposure and low sensitivity;

[0012] Step S12: Without changing the device sensitivity, extend the exposure time to obtain high-quality color images under normal light conditions, that is, normal-light color pictures, and use them as Ground Truth;

[0013] Step S13: Switch to the monochromatic camera, and set the sensitivity ratio of the monochromatic camera to the color camera to 5:1, and capture high-quality low-light monochromatic images at the same short exposure time as the color camera in step S11;

[0014] Step S14: Capture n pairs of data, and each pair of data includes low-light color pictures, low-light monochromatic pictures, and normal-light color pictures in the same scene. Preliminarily process the pictures in the paired low-light dataset and divide them into a training set and a test set.

[0015] Further, in step S3, in the given low-light color image I low The low-high frequency decoupling is used to extract frequency domain features of the low-light color image I by a perturbation corrector based on wavelet transform low And a lossless and reversible wavelet transform is used to construct a perturbation correction codec to efficiently reconstruct the degraded area, generating a perturbation correction map While preventing the loss of detail information in the aerial map, the whole process is described as:

[0016]

[0017] where θ p represents parameters of.

[0018] Further, in step S4, the corrected image obtained in step S3, that is, the perturbation correction map and the low-light monochromatic image I mono are used as the input of the refinement restorer based on monochromatic image guidance to Perform RGB to YUV color space conversion to process the luminance and chrominance branches separately, and introduce a low-light monochrome image in the Y branch to guide the refinement and restoration to obtain a high-quality aerial image I with good exposure and rich texture details en , as shown in Equation (2):

[0019]

[0020] where θ R represents parameters of.

[0021] Furthermore, the wavelet transform-based perturbation corrector adopts a three-scale U-shaped structure. The wavelet transform-based perturbation corrector processes the low-light color image I low to generate a perturbation correction map The specific implementation method is as follows:

[0022] The low-light color image I low is decomposed into a low-frequency global component and three directional high-frequency detail components (H, W, and C represent the image height, image width, and number of feature channels respectively) through lossless and reversible downsampling using the discrete wavelet transform DWT. The DWT decomposition process is expressed as:

[0023]

[0024] where DWT(·) represents the discrete wavelet transform;

[0025] The low-frequency global component and the three directional high-frequency detail components are processed by a three-layer wavelet sampling and degradation correction convolution module DRCM to extract low-frequency hierarchical features and high-frequency hierarchical features where i = 1, 2, 3;

[0026] The low-frequency hierarchical features are processed by the degradation correction convolution module DRCM to extract global information and correct uneven exposure. The high-frequency hierarchical features are fused through an adaptive feature fusion module AFFM. After integrating the high-frequency components in each direction and reducing the channel dimension from 3C to C, they are efficiently reconstructed through the degradation correction convolution module DRCM;

[0027] Finally, a perturbation correction low-frequency component and a perturbation correction high-frequency component are reconstructed through the inverse wavelet transform IWT and combined with a subtraction operation to obtain the perturbation correction map As shown in formula (4):

[0028]

[0029] Where IWT(·) represents the inverse wavelet transform.

[0030] Furthermore, the degradation correction convolution module DRCM consists of two layers of layer normalization LN, the coordinate attention convolution block CACM, and the feed-forward network FFN. The specific representation formula of the degradation correction convolution module DRCM is as follows:

[0031]

[0032] Where, F in , F att , F out are the input feature, the attention feature, and the output feature respectively; CACM realizes the extraction of global image information. First, it extracts the feature map through the convolutional layer, decomposes the channel attention into two orthogonal one-dimensional feature codes, generates two attention maps in the horizontal and vertical directions, and is used to capture the long-range dependence of the input feature map along the spatial direction;

[0033] Y att = Y pool (f input ), X att = X pool (f input ), (6)

[0034] Where, X pool (·) and Y pool (·) are the horizontal and vertical pooling operations respectively, X att (·) and Y att (·) are the horizontal and vertical attention maps respectively; multiply the attention maps in the two directions by the input feature to emphasize the representation of the attention area:

[0035] f output = Conv(Y att )·Conv(X att )·f input (7)

[0036] Where, f input and f output are the input and output features respectively, and Conv(·) represents the convolution operation. Furthermore, the refinement restorer R(·) based on the monochromatic image guidance adopts a double-branch three-scale U-shaped structure to refine and restore the corrected image, generating a high-quality aerial image with good exposure and rich texture details. The specific implementation method is:

[0037] Using the color space conversion from RGB to YUV, the perturbation correction map Decompose into luminance components and chrominance components and perform encoding processing on different branches respectively;

[0038] Process the UV branch features, chrominance components Through the Degradation Correction Convolution Module (DRCM) and the Degradation Correction Transformer Module (DRTM), suppress color noise and color deviation, and extract high-quality chrominance features

[0039] Process the Y branch luminance features, extract high-quality monochromatic image features from the low-light monochromatic image I mono Extract high-quality monochromatic image features Guide the luminance components Dynamically correct exposure and restore details, and sequentially pass through the Reference-based Illumination Fusion Module (RIFM), the Degradation Correction Convolution Module (DRCM), the Degradation Correction Transformer Module (DRTM), and the Illumination Detail Transformation Module (IDTM) to optimize exposure uniformity and restore more detail information, generating high-quality luminance features

[0040] Align and fuse the high-quality luminance features and the high-quality chrominance features Perform alignment fusion and generate a fine restoration map through the color space conversion from YUV to RGB Multiply with the perturbation correction map to obtain a high-quality aerial image I with good exposure and rich details en , achieving luminance balance, color restoration, and detail enhancement

[0041] Furthermore, for the Reference-based Illumination Fusion Module (RIFM), introduce a reference low-light monochromatic image and adaptively align the input image with the reference low-light monochromatic image using modulated deformable convolution; specifically, process the input feature F in and the reference low-light monochromatic image feature F m through the offset convolution layer to generate an offset Δf and a mask Δm, producing an aligned feature where C is the number of feature channels, and H×W is the spatial size; finally, propose an Alignment Reference Integration Module (ARIM) to further integrate and F m , effectively fusing complementary information into the enhanced aligned feature The overall process is written in the following form:

[0042]

[0043] where, M def (·) and M off (·) represent deformable convolution and offset convolution respectively

[0044] Furthermore, the Degradation Correction Transformer Module (DRTM) utilizes the reference low-light monochromatic image F m to restore overexposed and underexposed areas. By integrating the Illumination Detail Conversion Module (IDTM) into the Transformer to construct the DRTM, it captures global information to map the exposure distribution and utilizes the luminance details in the low-light monochromatic image to balance the exposure and enhance spatial details. The whole process is described as follows:

[0045]

[0046] where MHSA(·) represents the multi-head self-attention, LN represents the normalization LN, FFN represents the feed-forward network, F in 、F att 、F out are the input feature, the attention feature, and the output feature respectively, and Q, K, V represent the query, key, and value projections in the DRTM multi-head self-attention mechanism respectively.

[0047] Furthermore, the Illumination Detail Conversion Module (IDTM) utilizes the well-exposed low-light monochromatic image to provide more descriptive features for the illumination component through a correlation matching scheme, enhancing the representation ability of the luminance information; specifically, it calculates the matrix similarity M between the input feature F in and the low-light monochromatic image feature F m along the channel dimension; secondly, it selects the Top-1 vector, sorts it to generate the index S, and uses it to retrieve the corresponding similar channels from F m as the output; finally, it concatenates the selected feature Y selected with the original luminance feature to generate the fused feature The gating mechanism combines the luminance features to generate the output F out .

[0048] Compared with the prior art, the present invention has the following beneficial effects: The present invention can accurately generate high-quality aerial images with good exposure and rich texture details. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the overall working flow chart in the embodiment of the present invention;

[0050] Figure 2 is the low-light aerial imaging model diagram based on a color monochromatic camera in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The technical solutions of the present invention will be specifically described below with reference to the accompanying drawings.

[0052] Please refer to Figure 1, the present invention provides a color - monochrome dual - camera UAV aerial imaging method for low - light conditions, including the following steps:

[0053] Step S1: Use the color camera and monochrome camera carried by the aerial photography device to obtain the required paired low - illumination data sets at low altitude at night;

[0054] Step S2: Construct the overall structure of the aerial low - light imaging method ALICE based on the color - monochrome cameras;

[0055] Step S3: Construct a perturbation corrector based on wavelet transform in ALICE to roughly correct the degraded areas of the low - light color images in the paired low - illumination data sets obtained in Step S1 to obtain corrected images, while avoiding the loss of unique texture information in the aerial photography scene;

[0056] Step S4: Construct a refinement restorer guided by the monochrome image in ALICE to refine and restore the corrected images obtained in Step S3, and use the low - light monochrome images in the paired low - illumination data sets to guide the fine restoration of exposure and details, generating high - quality aerial images with good exposure and rich texture details.

[0057] In this embodiment, Step S1 is specifically as follows:

[0058] Step S11: In a low - altitude environment at night, use the aerial photography device to collect low - light color images with short exposure and low ISO to minimize motion blur and photosensitive noise as much as possible;

[0059] Step S12: While keeping the device's ISO unchanged, extend the exposure time to obtain high - quality color images under normal lighting conditions, and use them as GroundTruth for subsequent model training and evaluation;

[0060] Step S13: Switch to the monochrome camera, and set the ISO ratio of the monochrome camera to the color camera to 5:1, and capture high - quality low - light monochrome images at the same short exposure time as the color camera for enhancing detail restoration and brightness correction;

[0061] Step S14: Capture n pairs of data, and each pair of data includes low - light color images, low - light monochrome images, and normal - light monochrome images in the same scene. Perform preliminary processing on the images in the data set and divide them into a training set and a test set.

[0062] In this embodiment, in a real nighttime environment, the sensitivity of the device lens is set to ISO: 160, and low-light color aerial images are collected under the condition that the exposure time is less than 0.1 s. Subsequently, on the premise of keeping the sensitivity unchanged, the exposure time is adjusted to the long-exposure mode to obtain normal-light color images as Ground Truth. In addition, to obtain high-quality low-light black-and-white aerial images, we switch to a monochrome camera, set its exposure time to be the same as that of the low-light color image, and set the sensitivity ratio of the monochrome camera to the color camera to 5:1 to ensure sufficient light signal capture.

[0063] In this embodiment, the model image data is specifically as follows: A total of 30,000 pairs of images with a resolution of 160×480 are collected by cropping the original data, and the training set and the validation set are divided according to a ratio of 9:1. To evaluate the generalization ability of the proposed method ALICE, we select image pairs with different resolutions in the dataset for evaluation, and crop 1,000 pairs of images with a resolution of 480×480 and 200 pairs of images with a resolution of 1440×960 as the test set to comprehensively evaluate the adaptability and performance of the model at different resolutions.

[0064] In this embodiment, referring to Figure 2 , the aerial low-light imaging model (ALICE) based on a color-monochrome camera adopts a coarse-to-fine strategy to correct aerial low-light degradation, that is, a perturbation corrector based on wavelet transform and a refinement restorer guided by a monochrome image construct an aerial low-light imaging model (ALICE) based on a color-monochrome camera, adopt a coarse-to-fine strategy to correct aerial low-light degradation, and generate high-quality aerial images; among them,

[0065] In a given low-light color image I low , through a perturbation corrector based on wavelet transform perform low-high frequency decoupling on the low-light color image I low to extract frequency domain features, and use a lossless and reversible wavelet transform to construct an encoder-decoder for perturbation correction to efficiently reconstruct the degraded area and generate a perturbation correction map while preventing the loss of detail information in the aerial image. The above process can be described as:

[0066]

[0067] where θ p represents 's parameters;

[0068] Then, the output corrected image, that is, the perturbation correction map and the low-light monochrome image I mono are used as the refinement restorer guided by a monochrome image The input, for perform RGB to YUV color space conversion, process the luminance and chrominance branches separately, and introduce a low-light monochrome image in the Y branch to guide the refinement and restoration to obtain an aerial image I with good exposure and rich texture details en , as shown in Equation (2):

[0069]

[0070] where θ R represents the parameter of

[0071] In this embodiment, referring to Figure 2 , in step S3, the perturbation corrector P(·) based on wavelet transform adopts a three-scale U-shaped structure and uses discrete wavelet transform (DWT) for lossless reversible downsampling to alleviate the content loss problem caused by traditional sampling methods and well retain the rich detail information in the aerial image, specifically as follows:

[0072] Step S31: The low-light color image I low is decomposed into a low-frequency global component and three directional high-frequency detail components through lossless reversible downsampling using discrete wavelet transform (DWT) to alleviate the content loss problem caused by traditional downsampling methods. The DWT decomposition process is expressed as:

[0073]

[0074] where DWT(·) represents discrete wavelet transform;

[0075] Step S32: The low-frequency global component and the three directional high-frequency detail components are processed by a three-layer wavelet sampling and degradation correction convolution module (DRCM) to extract low-frequency hierarchical features and high-frequency hierarchical features where i = 1, 2, 3;

[0076] Step S33: The low-frequency hierarchical feature is processed by DRCM to extract global information and correct uneven exposure, while the high-frequency hierarchical feature is fused through an adaptive feature fusion module (AFFM), integrates the directional high-frequency components, reduces the channel dimension from 3C to C, and then passes through DRCM for efficient reconstruction;

[0077] Step S34: Finally, generate the perturbation-corrected low-frequency component and the perturbation-corrected high-frequency component Reconstruction is performed through inverse wavelet transform (IWT), combined with subtraction operation, and finally a perturbation correction map is obtained. As shown in formula (4):

[0078]

[0079] Where IWT(·) represents the inverse wavelet transform.

[0080] In this embodiment, the degradation correction convolution module (DRCM) efficiently extracts the global information of the image, thereby accurately locating the degradation area of the image and improving the problem of uneven exposure distribution in low-light aerial images. DRCM consists of two layers of normalization (LN), a coordinate attention convolution block (CACM), and a feed-forward network (FFN). The whole process can be expressed as:

[0081]

[0082] Where F in , F att , F out are the input feature, the attention feature, and the output feature. CACM extracts the global information of the image with relatively low computational complexity. First, feature maps are extracted through convolutional layers. The channel attention is decomposed into two orthogonal one-dimensional feature codes, generating two attention maps in the horizontal and vertical directions to capture the long-range dependencies of the input feature map along the spatial direction.

[0083] Y att = Y pool (f input ), X att = X pool (f input ), (6)

[0084] Where X pool (·) and Y pool (·) are the horizontal and vertical pooling operations respectively. X att (·) and Y att (·) are the horizontal and vertical attention maps respectively. Finally, we multiply the attention maps in the two directions by the input feature to emphasize the representation of the attention area:

[0085] f output = Conv(Y att )·Conv(X att )·f input . (7)

[0086] In this embodiment, referring to Figure 2 , in step S4, the monochromatic image-guided refinement restorer Adopt a dual-branch three-scale U-shaped structure, utilize the color space conversion from RGB to YUV, and decompose the perturbation correction map into the luminance component and the chrominance component and perform encoding processing on different branches respectively; specifically as follows:

[0087] Step S41, process the UV branch features. The chrominance component passes through the Degradation Rectification Convolution Module (DRCM) and the Degradation Rectification Transformer Module (DRTM) to effectively suppress color noise and color deviation and extract high-quality chrominance features

[0088] Step S42, process the Y branch luminance features. Extract high-quality monochrome image features from the low-light monochrome image I mono to guide the luminance component to dynamically correct exposure and restore details, and sequentially pass through the Reference-based Illumination Fusion Module (RIFM), the Degradation Rectification Convolution Module (DRCM), the Degradation Rectification Transformer Module (DRTM), and the Illumination Detail Conversion Module (IDTM) to optimize exposure uniformity and restore more detail information to generate high-quality luminance features

[0089] Step S43, align and fuse the luminance features in Step S42 and the chrominance features in Step S41 and perform color space conversion from YUV to RGB to generate a fine restoration map Multiply it with the perturbation correction map to obtain a well-exposed and detail-rich aerial image I en , achieving luminance balance, color restoration, and detail enhancement.

[0090] In this embodiment, the Reference-based Illumination Fusion Module (RIFM) introduced the reference low-light monochrome image and adaptively aligned the input image with the reference low-light monochrome image using modulated deformable convolution. Specifically, the input feature F in and the reference low-light monochrome image feature F m are processed through the offset convolution layer to generate the offset Δf and the mask Δm, generating the aligned feature Finally, an Aligned Reference Integration Module (ARIM) is proposed to further integrate and F m , effectively fusing complementary information into the enhanced aligned feature It can be written in the following form:

[0091] ​

[0092] Among them, M def (·) and M off (·) respectively represent deformable convolution and offset convolution, and ARIM(·) represents an alignment reference integration module.

[0093] In this embodiment, the Degradation Rectification Transformer Module (DRTM) utilizes the low-light monochromatic image feature F m to effectively restore overexposed and underexposed regions. By integrating the Illumination Detail Transformation Module (IDTM) into the Transformer to construct the Degradation Rectification Transformer Module (DRTM), global information is captured to map the exposure distribution, and the luminance details in the low-light monochromatic image are used to balance the exposure and enhance the spatial details. The above process is expressed as:

[0094]

[0095] Among them, MHSA(·) represents multi-head self-attention.

[0096] In this embodiment, the Illumination Detail Transformation Module (IDTM) uses a well-exposed monochromatic image to provide more descriptive features for the illumination component through a correlation matching scheme, enhancing the representation ability of the luminance information. Specifically, the matrix similarity M between the input feature F in and the monochromatic image feature F m is calculated along the channel dimension. Secondly, the Top-1 vector is selected, sorted to generate the index S, and used to retrieve the corresponding similar channels from F m as the output. Finally, the selected feature Y selected is concatenated with the monochromatic image feature F m to generate a fused feature The gating mechanism combines the luminance features to generate the output F out . The above process is expressed as:

[0097] M = CMS(F in , F m ),

[0098] S = Sort(Top1(M)),

[0099] Y selected = Select(F m |S),

[0100]

[0101] The above are the preferred embodiments of the present invention. All changes made according to the technical solutions of the present invention that do not exceed the scope of the technical solutions of the present invention in terms of the functions and effects produced belong to the protection scope of the present invention.

Claims

1. A color - monochrome dual - camera UAV aerial photography imaging method for low - light conditions, characterized in that, It includes the following steps: Step S1: Use the color camera and monochrome camera carried by the aerial photography equipment to obtain the required paired low-illumination datasets at low altitude at night; Step S2: Construct the overall structure of the aerial low-light imaging method ALICE based on the color-monochrome camera; Step S3: Construct a perturbation corrector based on wavelet transform in ALICE to roughly correct the degraded areas of the low-light color images in the paired low-illumination datasets obtained in Step S1 to obtain corrected images, while avoiding the loss of texture information unique to the aerial photography scene; Step S4: Construct a refinement restorer guided by the monochrome image in ALICE to refine and restore the corrected images obtained in Step S3, and use the low-light monochrome images in the paired low-illumination datasets to guide the fine restoration of exposure and details to generate high-quality aerial photography images with good exposure and rich texture details.

2. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 1, wherein, Specifically, Step S1 is as follows: Step S11: In the low-altitude environment at night, use the color camera carried by the aerial photography equipment to collect low-light color images with short exposure and low ISO; Step S12: Without changing the ISO of the equipment, extend the exposure time to obtain high-quality color images under normal lighting conditions, that is, normal-light color images, and use them as Ground Truth; Step S13: Switch to the monochrome camera, set the ISO ratio of the monochrome camera to the color camera to 5:1, and capture high-quality low-light monochrome images at the same short exposure time as the color camera in Step S11; Step S14: Capture n pairs of data, and each pair of data includes low-light color images, low-light monochrome images, and normal-light color images in the same scene. Preliminarily process the images in the paired low-illumination datasets and divide them into a training set and a test set.

3. The method for aerial photography imaging of a color-monochrome dual-camera drone facing low light according to claim 1, wherein, In step S3, in the given low-light color image I low a low-high frequency decoupling is performed by a wavelet transform-based perturbation corrector on the low-light color image I low to extract frequency domain features, and a lossless and reversible wavelet transform is used to construct a perturbation correction codec to efficiently reconstruct the degraded region, generating a perturbation correction map Meanwhile, to prevent the loss of detail information in the aerial map, the whole process is described as follows: where θ p represents parameters of 4. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 2, wherein In step S4, the corrected image obtained in step S3, i.e., the disturbance correction map and the low-light monochrome image I mono are used as the input of the monochrome image-guided refinement restorator . For , RGB to YUV color space conversion is performed to process the luminance and chrominance branches respectively, and the low-light monochrome image-guided refinement is introduced in the Y branch to obtain a high-quality aerial image I with good exposure and rich texture details en , as shown in formula (2): where θ R represents parameter.

5. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 3, wherein, Wavelet Transform-based Disturbance Corrector A wavelet transform-based disturbance corrector with a three-scale U-shaped structure Processes the low-light color image I low To generate a disturbance correction map The specific implementation method is as follows: The low-light color image I low is losslessly and reversibly downsampled by using the discrete wavelet transform (DWT) and decomposed into a low-frequency global component and three directional high-frequency detail components Let H, W, and C denote the image height, image width, and the number of feature channels respectively. The DWT decomposition process is expressed as: Where DWT(·) represents the discrete wavelet transform; Low-frequency global component and three-direction high-frequency detail components After being processed by the three-layer wavelet sampling and degradation correction convolution module DRCM, low-frequency hierarchical features are extracted and high-frequency hierarchical features where i = 1, 2, 3; Low-frequency layering features After being processed by the Degradation Correction Convolution Module (DRCM), global information is extracted and uneven exposure is corrected, resulting in high-frequency layering features Through the Adaptive Feature Fusion Module (AFFM), the high-frequency components in all directions are integrated and the channel dimension is reduced from 3C to C. After that, efficient reconstruction is achieved through the Degradation Correction Convolution Module (DRCM); Finally generate the low-frequency component of disturbance correction and the high-frequency component of disturbance correction Reconstruct through the inverse wavelet transform IWT and combine with subtraction operation to obtain the disturbance correction map As shown in formula (4): Where IWT(·) represents the inverse wavelet transform.

6. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 5, wherein The degradation correction convolutional module DRCM consists of two layers of layer normalization LN, a coordinate attention convolutional block CACM, and a feed-forward network FFN. The specific representation formula of the degradation correction convolutional module DRCM is as follows: Among them, F in , F att , F out are the input feature, the attention feature, and the output feature respectively; CACM realizes the extraction of global image information. First, it extracts the feature map through the convolutional layer, decomposes the channel attention into two orthogonal one-dimensional feature codes, generates two attention maps in the horizontal and vertical directions, and is used to capture the long-range dependence of the input feature map along the spatial direction; Y att = Y pool (f input ), X att = X pool (f input ), (6) Among them, X pool (·) and Y pool (·) are horizontal and vertical pooling operations respectively, and X att (·) and Y att (·) are horizontal and vertical attention maps respectively; multiply the attention maps in both directions by the input features to emphasize the representation of the attention regions: f output = Conv(Y att )·Conv(X att )·f input , (7) Among them, f input and f output are the input and output features respectively, and Conv(·) represents the convolution operation.

7. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 4, wherein Monochrome Image-Guided Refinement Restorer Adopting a double-branch three-scale U-shaped structure, it refines and restores the corrected image to generate a high-quality aerial image with good exposure and rich texture details. The specific implementation method is as follows: Using the RGB-to-YUV color space conversion, the disturbance-corrected image is decomposed into a luminance component and a chrominance component and encoding processes are performed on different branches respectively; Process the UV branch feature, chrominance component Through the Degradation Rectification Convolution Module (DRCM) and the Degradation Rectification Transformer Module (DRTM), color noise and color cast are suppressed, and high-quality chrominance features are extracted Process the brightness features of the Y-branch, starting from the low-light monochromatic image I mono Extract high-quality monochromatic image features Guide the luminance component Dynamically correct the exposure and restore details. Then, successively pass through the reference-based illumination fusion module RIFM, the degradation correction convolutional module DRCM, the degradation correction transformer module DRTM, and the illumination detail conversion module IDTM to optimize the exposure uniformity and restore more detail information, generating high-quality brightness features Align and fuse high-quality luminance features and high-quality chrominance features to generate a fine recovery mapping through the color space conversion from YUV to RGB Multiply it with the perturbation correction map to obtain a high-quality aerial image I with good exposure and rich details en , achieving luminance balance, color restoration, and detail enhancement.

8. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 7, wherein, Reference-based Illumination Fusion Module (RIFM), which introduces a reference low-light monochromatic image and adaptively aligns the input image with the reference low-light monochromatic image by using modulated deformable convolution; specifically, input feature F in and the reference low-light monochromatic image feature F m are processed through an offset convolutional layer to generate an offset Δf and a mask Δm, producing an aligned feature where C is the number of feature channels and H×W is the spatial size; Finally, an Alignment Reference Integration Module (ARIM) is proposed to further integrate and F m to effectively fuse complementary information into the enhanced alignment features The overall process is written in the following form: Among them, M def (·) and M off (·) represent deformable convolution and offset convolution respectively.

9. The method for aerial photography imaging of a color - monochrome dual - camera drone facing low - light according to claim 7, wherein, The Degradation Correction Transformer Module DRTM utilizes the reference low-light monochromatic image F m to restore overexposed and underexposed regions. By integrating the Illumination Detail Conversion Module IDTM into the Transformer to construct the Degradation Correction Transformer Module DRTM, it captures global information to map the exposure distribution and utilizes the luminance details in the low-light monochromatic image to balance the exposure and enhance the spatial details. The whole process is described as follows: Among them, MHSA(·) represents multi-head self-attention, LN represents layer normalization LN, FFN represents feed-forward network, F in 、F att 、F out are the input feature, attention feature, and output feature respectively, and Q, K, and V represent the query, key, and value projections in the DRTM multi-head self-attention mechanism respectively.

10. The method for aerial imaging of a color - monochrome dual - camera drone facing low - light as claimed in claim 7 or 9, wherein, The Illumination Detail Transformation Module (IDTM) uses well-exposed low-light monochromatic images to provide more descriptive features for the illumination component through a correlation matching scheme, enhancing the representation ability of brightness information. Specifically, the input feature F is calculated along the channel dimension in and the low-light monochromatic image feature F m to obtain the matrix similarity M therebetween. Secondly, the Top-1 vector is selected, sorted to generate the index S, and used to retrieve the corresponding similar channel from F m as the output. Finally, the selected feature Y selected is concatenated with the original brightness feature to generate the fused feature The gating mechanism combines the brightness features to generate the output F out .

Images