Image enhancement method and system based on sparsity under low-illumination conditions

Abstract

This invention discloses an image enhancement method and system based on sparsity under low-light conditions, belonging to the field of image processing technology. Considering the sparsity of images, this invention combines an attention mechanism and a feedforward neural network to construct a low-light drone image enhancement model. This model can effectively extract features and details from drone images under low-light conditions, solving the problems of color distortion or lack of detail in drone images, improving the image quality and clarity of the enhanced drone images, which is beneficial for drone identification, detection, or monitoring, and assists in the further utilization of drones. The system structure of this invention is simple, constructing a corresponding low-light drone image enhancement model to effectively enhance drone images under low-light conditions and improve the quality of drone nighttime images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image enhancement method and system based on sparsity under low illumination conditions. Background Technology

[0002] With the continuous development of drone technology, nighttime image enhancement has become a significant challenge in drone vision applications. In many drone missions, such as nighttime patrols, security patrols, and nighttime search and rescue, drones typically need to capture images under low-light conditions. This leads to a decline in image quality, resulting in problems such as low contrast, color distortion, noise amplification, and loss of detail, thus limiting the effectiveness and safety of drone operations.

[0003] Low-light image enhancement technology has broad application potential in the field of unmanned aerial vehicles (UAVs), improving the quality of nighttime images and enhancing the efficiency and accuracy of mission execution. Its applications extend beyond nighttime aerial photography to include: Nighttime UAV surveillance: UAVs used for nighttime surveillance and patrols can benefit from low-light image enhancement to improve image quality and increase the reliability of surveillance systems. Nighttime search and rescue: In nighttime search and rescue missions, UAVs can utilize low-light image enhancement to improve visibility in the search area, increasing the success rate of rescue operations. Nighttime military applications: UAVs used for nighttime military reconnaissance and surveillance can enhance images transmitted by UAVs, providing more intelligence and situation reports.

[0004] To address the problem of image enhancement under low-light conditions, previous methods mainly relied on traditional image processing techniques such as histogram equalization, filtering, and exposure enhancement. However, these methods often introduce additional noise or lose image detail when processing images under low-light conditions, thus limiting the enhancement effect. In recent years, deep learning-based image enhancement methods have made significant progress and have become the mainstream technology for low-light image enhancement. However, under low-light conditions, the information in the image is often affected by severe illumination inhomogeneity and noise interference. Therefore, an innovative method is needed to handle the image enhancement problem of UAVs under low-light conditions to improve image quality and preserve detail information. However, existing deep learning-based image enhancement methods do not consider the sparsity of low-light images, leading to problems such as color distortion or inadequate detail recovery in the enhanced image. Summary of the Invention

[0005] The purpose of this invention is to provide an image enhancement method and system based on sparsity under low-light conditions to improve the above-mentioned technical problems.

[0006] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0007] An image enhancement method based on sparsity under low-light conditions includes:

[0008] S1. Obtain the low-light drone image to be enhanced; the low-light drone image is an image taken by the drone under low-light conditions;

[0009] S2. Preprocess the low-light drone images to be enhanced;

[0010] S3. Based on image sparsity, construct a low-light UAV image enhancement model;

[0011] S4. Input the preprocessed low-light drone image into the low-light drone image enhancement model, and output the enhanced drone image.

[0012] Furthermore, the preprocessing of the low-light drone image to be enhanced includes:

[0013] S2-1. Randomly crop the low-light drone image to be enhanced to obtain the cropped low-light drone image.

[0014] S2-2. Perform data augmentation on the cropped low-light drone image to obtain a preprocessed low-light drone image; the data augmentation includes random flipping, rotation, and random cropping.

[0015] Furthermore, the low-light UAV image enhancement model includes a cascaded background modeling-detail restoration sub-model, a global feature fusion module, and a Conv3 layer;

[0016] The background modeling-detail restoration sub-model comprises seven background modeling-detail restoration modules, designated as the first through seventh sub-modules. The outputs of the first sub-module are input to the second and seventh sub-modules, respectively; the outputs of the second sub-module are input to the third and sixth sub-modules, respectively; the outputs of the third sub-module are input to the fourth and fifth sub-modules, respectively; the outputs of the fourth sub-module are input to the fifth sub-module and the global feature fusion module, respectively; the outputs of the fifth sub-module are input to the sixth sub-module and the global feature fusion module, respectively; the outputs of the sixth sub-module are input to the seventh sub-module and the global feature fusion module, respectively; and the output of the seventh sub-module is input to the global feature fusion module.

[0017] The global feature fusion module consists of an upsampling layer, a first convolutional layer, a second convolutional layer, a first sparse self-attention layer, and a third convolutional layer connected in series. The first and third convolutional layers both use 1×1 convolutional kernels, while the second convolutional layer uses a 3×3 convolutional kernel.

[0018] Furthermore, each background modeling-detail recovery submodule includes a position encoding submodule, a convolutional-self-attention submodule, and a feedforward neural network submodule, all connected in series. The position encoding submodule includes a fourth convolutional layer connected in series and a DWConv layer with a 3×3 kernel. The convolutional-self-attention submodule includes parallel dilated convolutional layers and a second sparse self-attention layer. The dilated convolutional layers include a series of dilated... The Conv3 layer and the first activation layer; both the first and second sparse self-attention layers employ the sparse self-attention mechanism; the feedforward neural network submodule includes a cascaded normalization layer, a fifth convolutional layer, a convolutional submodule, and a sixth convolutional layer; the convolutional submodule includes a parallel seventh convolutional layer, a convolutional layer, and a ninth convolutional layer; the convolutional layer includes a cascaded eighth convolutional layer and a second activation layer; both the first and second activation layers employ the GELU activation function; the fourth, fifth, and sixth convolutional layers all employ 1×1 convolutional kernels; the seventh, eighth, and ninth convolutional layers all employ 3×3 convolutional kernels.

[0019] Furthermore, the training process of the low-light UAV image enhancement model includes:

[0020] S3-1. Obtain low-light drone training images and their labels. Use the same method as in S2 to preprocess the low-light drone training images to obtain preprocessed low-light drone training images.

[0021] S3-2. Input the preprocessed low-light UAV training images and their labels into the background modeling-detail restoration sub-model, and output the first training feature map data, the second training feature map data, the third training feature map data and the fourth training feature map data;

[0022] S3-3. Input the first training feature map data, the second training feature map data, the third training feature map data, and the fourth training feature map data into the global feature fusion module, and output the fused training feature map data.

[0023] S3-4. Input the fused training feature map data into the Conv3 layer and output the convolutional fused training feature map data.

[0024] S3-5. Add the convolutional fused training feature map data and the preprocessed low-light drone training image to output the enhanced drone training image.

[0025] S3-6. Based on the enhanced UAV training images, construct the corresponding training loss function;

[0026] S3-7. Based on the training loss function, the gradient of the low-light UAV image enhancement model is calculated using the backpropagation method, and the parameters of the low-light UAV image enhancement model are adjusted using the Adam lake-hopping algorithm.

[0027] S3-8. Repeat S3-1 to S3-7 until the training error is less than the error threshold or the preset number of iterations is reached to complete the training of the low-light drone image enhancement model.

[0028] Further, S3-2 includes the following steps:

[0029] S3-2-1. Input the preprocessed low-light UAV training image into the first background modeling-detail restoration submodule and output the first initial training feature map data.

[0030] S3-2-2: Input the first initial training feature map data into the second background modeling-detail restoration submodule, and output the second initial training feature map data.

[0031] S3-2-3. Input the second initial training feature map data into the third background modeling-detail restoration submodule, and output the third initial training feature map data.

[0032] S3-2-4. Input the third initial training feature map data into the fourth background modeling-detail restoration submodule, and output the first training feature map data.

[0033] S3-2-5. Input the third initial training feature map data and the first training feature map data into the fifth background modeling-detail restoration submodule, and output the second training feature map data.

[0034] S3-2-6. Input the second initial training feature map data and the second training feature map data into the sixth background modeling-detail restoration submodule, and output the third training feature map data.

[0035] S3-2-7. Input the first initial training feature map data and the third training feature map data into the seventh background modeling-detail restoration submodule, and output the fourth training feature map data.

[0036] Further, S3-2-5 includes the following steps:

[0037] S3-2-5-1. Concatenate the third initial training feature map data and the first training feature map data to obtain the concatenated initial training feature map data.

[0038] S3-2-5-2: Input the stitched initial training feature map data into the position encoding submodule of the fifth background modeling-detail restoration submodule, and output the encoded initial training feature map data.

[0039] S3-2-5-3. Input the encoded initial training feature map data into the convolutional-self-attention submodule of the fifth background modeling-detail recovery submodule, and output the initial training key feature map data.

[0040] S3-2-5-4. Input the initial training key feature map data into the feedforward neural network submodule of the fifth background modeling-detail recovery submodule, and output the second training feature map data.

[0041] Furthermore, S3-3 includes the following steps:

[0042] S3-3-1, Use the first training feature map data, the second training feature map data, the third training feature map data, and the fourth training feature map data as training feature map data. ;in, Indicates the sequence number;

[0043] S3-3-2, Transfer the training feature map data The input is fed into the upsampling layer, and the output is the sampled training feature map data;

[0044] S3-3-3: Input the sampled training feature map data into the first convolutional layer, and output the training feature map data with unified dimensions.

[0045] S3-3-4. Input the training feature map data after unifying the dimensions into the second convolutional layer, and output the encoded training feature map data.

[0046] S3-3-5. Input the encoded training feature map data into the first sparse self-attention layer and output the enhanced training feature map data.

[0047] S3-3-6. Input the enhanced training feature map data into the third convolutional layer and output the fused training feature map data.

[0048] A sparsity-based image enhancement system for low-light conditions includes:

[0049] The drone image acquisition module is used to acquire low-light drone images that need to be enhanced.

[0050] The image preprocessing module is used to preprocess the low-light drone images to be enhanced;

[0051] The low-light drone image enhancement module is used to construct a low-light drone image enhancement model based on image sparsity. The preprocessed low-light drone image is input into the low-light drone image enhancement model, and the enhanced drone image is output.

[0052] The beneficial effects of this invention are as follows:

[0053] This method considers the sparsity of images and combines attention mechanisms and feedforward neural networks to construct a low-light drone image enhancement model. It can effectively extract features and details of drone images under low-light conditions, solve the problems of color distortion or lack of details in drone images, improve the image quality and clarity of enhanced drone images, and facilitate the identification, detection or monitoring of drones, thus assisting in the further utilization of drones.

[0054] This system has a simple structure and constructs a corresponding low-light drone image enhancement model to effectively enhance drone images under low-light conditions and improve the quality of drone nighttime images. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the background modeling-detail restoration sub-model in an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the global feature fusion module in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram of the background modeling module in an embodiment of the present invention;

[0060] Figure 5 This is a system structure diagram in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 This embodiment provides an image enhancement method based on sparsity under low-light conditions, which includes:

[0063] S1. Acquire low-light drone images to be enhanced;

[0064] S2. Preprocess the low-light drone images to be enhanced;

[0065] The preprocessing of the low-light drone images to be enhanced includes:

[0066] S2-1. Randomly crop the low-light drone images to be enhanced, using images of all sizes. The low-light drone image to be enhanced is randomly cropped to obtain a size of [size missing]. The drone image, i.e., the cropped low-light drone image; where, Indicates the altitude of a low-light drone image. This indicates the width of the low-light drone image. This indicates the height and width of the cropped low-light drone image;

[0067] S2-2. Perform data augmentation on the cropped low-light drone image to obtain a preprocessed low-light drone image; the data augmentation includes random flipping, rotation, and random cropping.

[0068] S3. Based on image sparsity, construct a low-light UAV image enhancement model;

[0069] like Figure 2 As shown, the low-light UAV image enhancement model includes a cascaded background modeling-detail restoration sub-model, a global feature fusion module, and a Conv3 layer;

[0070] The background modeling-detail restoration sub-model comprises seven background modeling-detail restoration modules, designated as the first through seventh sub-modules. The outputs of the first sub-module are input to the second and seventh sub-modules, respectively; the outputs of the second sub-module are input to the third and sixth sub-modules, respectively; the outputs of the third sub-module are input to the fourth and fifth sub-modules, respectively; the outputs of the fourth sub-module are input to the fifth sub-module and the global feature fusion module, respectively; the outputs of the fifth sub-module are input to the sixth sub-module and the global feature fusion module, respectively; the outputs of the sixth sub-module are input to the seventh sub-module and the global feature fusion module, respectively; and the output of the seventh sub-module is input to the global feature fusion module.

[0071] like Figure 3 As shown, the global feature fusion module includes an upsampling layer, a first convolutional layer, a second convolutional layer, a first sparse self-attention layer, and a third convolutional layer connected in series. The first and third convolutional layers both use 1×1 convolutional kernels, while the second convolutional layer uses a 3×3 convolutional kernel.

[0072] like Figure 4 As shown, each background modeling-detail recovery submodule includes a location encoding submodule, a convolutional-self-attention submodule, and a feedforward neural network submodule, all connected in series. The location encoding submodule includes a fourth convolutional layer and a 3×3DWConv layer connected in series. The convolutional-self-attention submodule includes parallel dilated convolutional layers and a second sparse self-attention layer. The dilated convolutional layers include a series of dilated... The Conv3 layer and the first activation layer; the first and second sparse self-attention layers both employ the sparse self-attention mechanism; the feedforward neural network submodule includes a cascaded normalization layer, a fifth convolutional layer, a convolutional submodule, and a sixth convolutional layer; the convolutional submodule includes a parallel seventh convolutional layer, a convolutional layer, and a ninth convolutional layer; the convolutional layer includes a cascaded eighth convolutional layer and a second activation layer; the first and second activation layers both employ the GELU activation function; the fourth, fifth, and sixth convolutional layers all employ 1×1 convolutional kernels; the seventh, eighth, and ninth convolutional layers all employ 3×3 convolutional kernels.

[0073] The training process of the low-light UAV image enhancement model includes:

[0074] S3-1. Obtain low-light drone training images and match them with corresponding labels; preprocess the low-light drone training images using the same method as in S2 to obtain preprocessed low-light drone training images; wherein, the label is the GroundTruth image corresponding to the low-light drone training image, and is processed in the same way as the low-light drone training image.

[0075] S3-2. Input the preprocessed low-light UAV training images and their labels into the background modeling-detail restoration sub-model, and output the first training feature map data, the second training feature map data, the third training feature map data and the fourth training feature map data;

[0076] S3-2 includes the following steps:

[0077] S3-2-1. Input the preprocessed low-light UAV training image into the first background modeling-detail restoration submodule and output the first initial training feature map data.

[0078] S3-2-2: Input the first initial training feature map data into the second background modeling-detail restoration submodule, and output the second initial training feature map data.

[0079] S3-2-3. Input the second initial training feature map data into the third background modeling-detail restoration submodule, and output the third initial training feature map data.

[0080] S3-2-4. Input the third initial training feature map data into the fourth background modeling-detail restoration submodule, and output the first training feature map data.

[0081] S3-2-5. Input the third initial training feature map data and the first training feature map data into the fifth background modeling-detail restoration submodule, and output the second training feature map data.

[0082] S3-2-6. Input the second initial training feature map data and the second training feature map data into the sixth background modeling-detail restoration submodule, and output the third training feature map data.

[0083] S3-2-7. Input the first initial training feature map data and the third training feature map data into the seventh background modeling-detail restoration submodule, and output the fourth training feature map data.

[0084] S3-2-5 includes the following steps:

[0085] S3-2-5-1. Concatenate the third initial training feature map data and the first training feature map data to obtain the concatenated initial training feature map data.

[0086] S3-2-5-2: Input the stitched initial training feature map data into the position encoding submodule of the fifth background modeling-detail restoration submodule, and output the encoded initial training feature map data. Increase the image dimension of the stitched initial training feature map data through a fourth convolutional layer with a 1×1 kernel, resulting in a size of... The processed initial training feature map data Position encoding is performed using a 3×3 DWConv layer (depth-separable convolution) to obtain the encoded initial training feature map data. .

[0087] The formula corresponding to S3-2-5-2 is:

[0088] ;

[0089] ;

[0090] in, This represents the initial training feature map data after concatenation. This indicates the fourth convolutional layer. This indicates the DWConv layer.

[0091] S3-2-5-3. Input the encoded initial training feature map data into the convolutional-self-attention submodule of the fifth background modeling-detail recovery submodule, and output the initial training key feature map data.

[0092] The formula for the convolutional self-attention submodule is:

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] in, , , These represent the query vector, key vector, and value vector, respectively. This represents the activation function. This indicates that the operation is divided equally according to the channel dimension. Indicates sparse self-attention. This represents the dimension of the key vector. This represents the output data of the second sparse self-attention layer. This represents the output data of the dilated convolutional layer. This indicates the dilated convolution operation corresponding to the Dilated Conv3 layer. This represents the activation function. The initial training key feature map data includes the output data of the second sparse self-attention layer. Output data of dilated convolutional layers .

[0098] S3-2-5-4. Input the initial training key feature map data into the feedforward neural network submodule of the fifth background modeling-detail recovery submodule, and the output is the second training feature map data. ;

[0099] Output data of the second sparse self-attention layer Output data of dilated convolutional layers The features are added together and then fed into the normalization layer. The dimensionality of the feature map output from the normalization layer is increased using a fifth convolutional layer with a 1×1 kernel, and then fed into the convolutional submodule. Within the convolutional submodule, three branches are formed: the first branch extracts features through a seventh convolutional layer with a 3×3 kernel; the second branch extracts features through an eighth convolutional layer with a 3×3 kernel, and then processes the feature map output from the eighth convolutional layer using the GELU activation function; the third branch extracts features through a ninth convolutional layer with a 3×3 kernel. The feature maps output from these three branches are then element-wise multiplied and then processed by a sixth convolutional layer with a 1×1 kernel to restore the original dimensionality. Therefore, the formula corresponding to S3-2-5-4 is:

[0100] ;

[0101] ;

[0102] ;

[0103] in, This represents the initial training key feature map data. This represents the concatenation function. Indicates an intermediate function. , These represent 1×1 convolution and 3×3 convolution, respectively. Indicates the normalization layer. This represents the activation function. Represents a 1×1 convolution. This represents the second training feature map data. This represents element-wise multiplication.

[0104] S3-2-6 to S3-2-7 adopt the same method as S3-2-5-1 to S3-2-5-4, and S3-2-1 to S3-2-4 adopt the same method as S3-2-2-2 to S3-2-2-4.

[0105] S3-3. Input the first training feature map data, the second training feature map data, the third training feature map data, and the fourth training feature map data into the global feature fusion module, and output the fused training feature map data.

[0106] S3-3 includes the following steps:

[0107] S3-3-1, Use the first training feature map data, the second training feature map data, the third training feature map data, and the fourth training feature map data as training feature map data. ;in, Indicates the sequence number;

[0108] S3-3-2. To unify the feature maps of different layers, the training feature map data will be... The input is fed into an upsampling layer for upsampling, and the output is the sampled training feature map data; when When the value is 1, the training feature map data is the first training feature map data, and so on. The numbers are 2, 3, and 4.

[0109] S3-3-3: Input the sampled training feature map data into the first convolutional layer with a 1×1 kernel to unify the dimensions, and output the training feature map data after dimension unification.

[0110] S3-3-4. Input the training feature map data after unifying the dimensions into the second convolutional layer with a 3×3 kernel for position encoding, and output the encoded training feature map data.

[0111] S3-3-5. Input the encoded training feature map data into the first sparse self-attention layer for feature enhancement, and output the enhanced training feature map data.

[0112] S3-3-6. Input the enhanced training feature map data into the third convolutional layer with a 1×1 kernel for convolution, and output the fused training feature map data. .

[0113] Therefore, the formula corresponding to S3-3 is:

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] in, This represents a 3×3 convolution. Indicates an intermediate function. Indicates upsampling, This represents the dimension of the key vector.

[0119] The global feature fusion module can fuse global information from preprocessed low-light drone training images, adaptively enhancing the background and detail features corresponding to the preprocessed low-light drone training images, thereby further improving the image enhancement capabilities of the low-light drone image enhancement model.

[0120] S3-4. Input the fused training feature map data into the Conv3 layer, and output the convolutional fused training feature map data. ;

[0121] S3-5. Add the convolutionally fused training feature map data and the preprocessed low-light UAV training image to obtain the enhanced UAV training image. The corresponding formula is:

[0122] ;

[0123] S3-6. Based on the enhanced UAV training images, construct the corresponding training loss function. The corresponding formula is:

[0124] ;

[0125] in, This represents the total number of enhanced drone training images. Represents low-light drone training images The corresponding Ground Truth image, Represents absolute value. This represents the summation function;

[0126] S3-7. Based on the training loss function, the gradient of the low-light UAV image enhancement model is calculated using the backpropagation method, and the parameters of the low-light UAV image enhancement model are adjusted using the Adam lake-hopping algorithm.

[0127] S3-8. Repeat S3-1 to S3-7 until the training error is less than the error threshold or the preset number of iterations is reached to complete the training of the low-light drone image enhancement model.

[0128] S4. Input the preprocessed low-light drone image into the low-light drone image enhancement model, and output the enhanced drone image.

[0129] like Figure 5 As shown, an image enhancement system based on sparsity under low-light conditions includes:

[0130] The drone image acquisition module is used to acquire low-light drone images that need to be enhanced.

[0131] The image preprocessing module is used to preprocess the low-light drone images to be enhanced;

[0132] The low-light drone image enhancement module is used to construct a low-light drone image enhancement model based on image sparsity. The preprocessed low-light drone image is input into the low-light drone image enhancement model, and the enhanced drone image is output.

[0133] In summary, the method of this invention considers the sparsity of images, combines attention mechanisms and feedforward neural networks to construct a low-light drone image enhancement model. This model can effectively extract features and details from drone images under low-light conditions, solve the problems of color distortion or lack of detail in drone images, and improve the image quality and clarity of the enhanced drone images. This is beneficial for drone identification, detection, or monitoring, and assists in the further utilization of drones. The system structure of this invention is simple, constructing a corresponding low-light drone image enhancement model to effectively enhance drone images under low-light conditions and improve the quality of drone nighttime images.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A sparse-based image enhancement method under low-illumination conditions, characterized in that, The method comprises the following steps: S1, obtaining a low-illumination unmanned aerial vehicle image to be enhanced; The low-illumination unmanned aerial vehicle image is an image taken by an unmanned aerial vehicle under low-illumination conditions; S2, preprocessing the low-illumination unmanned aerial vehicle image to be enhanced; S3, constructing a low-illumination unmanned aerial vehicle image enhancement model based on image sparsity; S4, inputting the preprocessed low-illumination unmanned aerial vehicle image into the low-illumination unmanned aerial vehicle image enhancement model to output an enhanced unmanned aerial vehicle image; The low-illumination unmanned aerial vehicle image enhancement model comprises a background modeling-detail restoration submodel, a global feature fusion module and a Conv3 layer connected in series; the background modeling-detail restoration submodel comprises seven background modeling-detail restoration submodules, which are respectively referred to as a first background modeling-detail restoration submodule to a seventh background modeling-detail restoration submodule; each background modeling-detail restoration submodule comprises a position coding submodule, a convolution-self-attention submodule and a feedforward neural network submodule connected in series; the convolution-self-attention submodule comprises a dilated convolution layer and a second sparse self-attention layer connected in parallel; the dilated convolution layer comprises a Dilated Conv3 layer and a first activation layer connected in series; The training process of the low-illumination unmanned aerial vehicle image enhancement model comprises the following steps: S3-1, obtaining a low-illumination unmanned aerial vehicle training image and a label thereof, preprocessing the low-illumination unmanned aerial vehicle training image by using the same method as that in S2 to obtain a preprocessed low-illumination unmanned aerial vehicle training image; S3-2, inputting the preprocessed low-illumination unmanned aerial vehicle training image and the label thereof into the background modeling-detail restoration submodel to output first training feature map data, second training feature map data, third training feature map data and fourth training feature map data; S3-3, inputting the first training feature map data, the second training feature map data, the third training feature map data and the fourth training feature map data into the global feature fusion module to output fused training feature map data; S3-4, inputting the fused training feature map data into the Conv3 layer to output convolutional fused training feature map data; S3-5, adding the convolutional fused training feature map data and the preprocessed low-illumination unmanned aerial vehicle training image to output an enhanced unmanned aerial vehicle training image; S3-6, constructing a corresponding training loss function based on the enhanced unmanned aerial vehicle training image; S3-7, calculating the gradient of the low-illumination unmanned aerial vehicle image enhancement model by using a back propagation method based on the training loss function, and adjusting the parameters of the low-illumination unmanned aerial vehicle image enhancement model by using an Adam optimization algorithm; S3-8, repeating S3-1 to S3-7 until the training error is less than an error threshold or a preset iteration number is reached, and the training of the low-illumination unmanned aerial vehicle image enhancement model is completed. 2.The sparse-based image enhancement method in low-illumination condition according to claim 1, wherein, The preprocessing of the low-illumination unmanned aerial vehicle image to be enhanced comprises the following steps: S2-1, randomly cutting the low-illumination unmanned aerial vehicle image to be enhanced to obtain a cut low-illumination unmanned aerial vehicle image; S2-2, data enhancement is performed on the cut low-illumination unmanned aerial vehicle image to obtain a preprocessed low-illumination unmanned aerial vehicle image; the data enhancement includes random flipping, rotation and random cropping. 3.The sparse-based image enhancement method in low-illumination condition according to claim 1, wherein, The global feature fusion module comprises, in sequence, an up-sampling layer, a first convolutional layer, a second convolutional layer, a first sparse self-attention layer and a third convolutional layer; the first convolutional layer and the third convolutional layer both adopt a 1×1 convolutional kernel; the second convolutional layer adopts a 3×3 convolutional kernel; The outputs of the first background modeling-detail restoration submodule are respectively input to the second background modeling-detail restoration submodule and the seventh background modeling-detail restoration submodule; the output of the second background modeling-detail restoration submodule is respectively input to the third background modeling-detail restoration submodule and the sixth background modeling-detail restoration submodule; the output of the third background modeling-detail restoration submodule is respectively input to the fourth background modeling-detail restoration submodule and the fifth background modeling-detail restoration submodule; the output of the fourth background modeling-detail restoration submodule is respectively input to the fifth background modeling-detail restoration submodule and the global feature fusion module; The output of the fifth background modeling-detail restoration submodule is respectively input to the sixth background modeling-detail restoration submodule and the global feature fusion module; The output of the sixth background modeling-detail restoration submodule is respectively input to the seventh background modeling-detail restoration submodule and the global feature fusion module; The output of the seventh background modeling-detail restoration submodule is input to the global feature fusion module. 4.The sparse-based image enhancement method in low-illumination condition according to claim 1, wherein, The position encoding submodule comprises, in sequence, a fourth convolutional layer and a DWConv layer with a 3×3 convolutional kernel; The first sparse self-attention layer and the second sparse self-attention layer both adopt a sparse self-attention mechanism; the feedforward neural network submodule comprises, in sequence, a normalization layer, a fifth convolutional layer, a convolutional submodule and a sixth convolutional layer; the convolutional submodule comprises, in parallel, a seventh convolutional layer, a convolutional layer and a ninth convolutional layer; the convolutional layer comprises, in sequence, an eighth convolutional layer and a second activation layer; the first activation layer and the second activation layer both adopt a GELU activation function; the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer all adopt a 1×1 convolutional kernel; the seventh convolutional layer, the eighth convolutional layer and the ninth convolutional layer all adopt a 3×3 convolutional kernel. 5.The sparse-based image enhancement method in low-illumination condition according to claim 4, wherein, The S3-2 comprises the following steps: S3-2-1, inputting the preprocessed low-illumination unmanned aerial vehicle training image to the first background modeling-detail restoration submodule to output first initial training feature map data; S3-2-2, inputting the first initial training feature map data to the second background modeling-detail restoration submodule to output second initial training feature map data; S3-2-3, inputting the second initial training feature map data to the third background modeling-detail restoration submodule to output third initial training feature map data; S3-2-4, inputting the third initial training feature map data to the fourth background modeling-detail restoration submodule to output first training feature map data; S3-2-5, input the third initial training feature map data and the first training feature map data into a fifth background modeling-detail restoration submodule, and output to obtain second training feature map data; S3-2-6, input the second initial training feature map data and the second training feature map data into a sixth background modeling-detail restoration submodule, and output to obtain third training feature map data; S3-2-7, input the first initial training feature map data and the third training feature map data into a seventh background modeling-detail restoration submodule, and output to obtain fourth training feature map data. 6.The sparse-based image enhancement method in low-illumination condition according to claim 5, wherein, The S3-2-5 includes the following steps: S3-2-5-1, splice the third initial training feature map data and the first training feature map data to obtain spliced initial training feature map data; S3-2-5-2, input the spliced initial training feature map data into a position coding submodule of the fifth background modeling-detail restoration submodule, and output to obtain coded initial training feature map data; S3-2-5-3, input the coded initial training feature map data into a convolution-self-attention submodule of the fifth background modeling-detail restoration submodule, and output to obtain initial training key feature map data; S3-2-5-4, input the initial training key feature map data into a feedforward neural network submodule of the fifth background modeling-detail restoration submodule, and output to obtain second training feature map data. 7.The sparse-based image enhancement method in low-illumination condition according to claim 4, wherein, The S3-3 includes the following steps: S3-3-1, taking the first training feature map data, the second training feature map data, the third training feature map data, and the fourth training feature map data as training feature map data ; wherein, denotes the arrangement serial number; S3-3-2, the training feature map data input to the up-sampling layer, and output the sampled training feature map data; S3-3-3, input the sampled training feature map data into a first convolutional layer, and output to obtain dimension-unified training feature map data; S3-3-4, input the dimension-unified training feature map data into a second convolutional layer, and output to obtain coded training feature map data; S3-3-5, input the coded training feature map data into a first sparse self-attention attention layer, and output to obtain enhanced training feature map data; S3-3-6, input the enhanced training feature map data into a third convolutional layer, and output to obtain fused training feature map data.

8. A sparse-based image enhancement system under low-light conditions for implementing a sparse-based image enhancement method under low-light conditions according to any one of claims 1 to 7, characterized in that, It includes: An unmanned aerial vehicle image acquisition module for acquiring a low-illumination unmanned aerial vehicle image to be enhanced; An image preprocessing module for preprocessing the low-illumination unmanned aerial vehicle image to be enhanced; A low-illumination unmanned aerial vehicle image enhancement module for constructing a low-illumination unmanned aerial vehicle image enhancement model based on image sparsity; Input the preprocessed low-illumination unmanned aerial vehicle image into the low-illumination unmanned aerial vehicle image enhancement model, and output to obtain an enhanced unmanned aerial vehicle image.

Images