Low-illumination image enhancement method based on wavelet-Fourier transform

Abstract

The invention discloses a low-illumination image enhancement method based on wavelet-Fourier transform. Relates to the technical field of computer vision, in particular to the technical field of low-illumination image enhancement based on wavelet-Fourier transform. The advantages of the two frequency domain transformations are fully exerted, the problems of insufficient illumination adjustment and poor detail retention in the existing method are solved, and the utilization efficiency of the frequency domain information is improved. The method comprises the following steps: acquiring an image data set, wherein the image data set comprises a low-illumination image and a real image; constructing an image enhancement model: inputting the low-illumination image into the image enhancement model, carrying out iterative training, and when the number of iterative training times reaches 200, obtaining a trained image enhancement model; and inputting the low-illumination image into the trained image enhancement model to obtain an enhanced image. Compared with the prior art, the method can better restore image details, adjust image saturation, reduce noise and improve visual effect.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to the field of low-light image enhancement technology based on wavelet-Fourier transform. Background Art

[0002] Images taken in low-light environments often suffer from problems such as low brightness, poor contrast, loss of details, and increased noise due to insufficient light, seriously affecting the visual effect of the images and the accuracy of subsequent computer vision tasks. Traditional methods such as histogram equalization and gamma correction can improve the image brightness to a certain extent, but when dealing with low-light images, they are prone to over-saturation in bright areas, loss of details in dark areas, and amplification of noise.

[0003] In recent years, the method of convolutional neural network (CNN) based on deep learning has gradually become the mainstream, which can automatically learn complex features and significantly improve the quality of low-light images. However, existing methods based on frequency domain analysis are mostly limited to single transformation or simple concatenation, failing to fully exploit the potential of wavelet transform and Fourier transform, resulting in low utilization efficiency of frequency domain information and problems such as uneven illumination, artifacts, and blurring in the restored images. Summary of the Invention

[0004] In view of the above problems, the present invention discloses a low-light image enhancement method based on wavelet-Fourier transform, which relates to the field of computer vision technology, gives full play to the advantages of the two frequency domain transformations, solves the problems of insufficient light adjustment and poor detail retention in existing methods, and improves the utilization efficiency of frequency domain information.

[0005] The method includes the following steps:

[0006] S1. Obtain an image data set, which includes: low-light image I input and real image I true ;

[0007] S2. Construct an image enhancement model: The image enhancement model sequentially passes through a YUV conversion module, three pre-denoising modules, three wavelet-Fourier enhancement modules, and a feature fusion module from input to output; the parameters of each module in the image enhancement model are updated through a loss function;

[0008] S21. Construct a YUV conversion module: The YUV conversion formula is: Y = 0.999R + 0.587G + 0.114B, U = -0.14713R - 0.28886G + 0.436B, V = 0.615R + 0.51499G + 0.10001B, where Y represents brightness, U represents chrominance, V represents saturation, R represents red, G represents green, and B represents blue;

[0009] The YUV conversion module converts the input low-light image from the RGB color space to the YUV color space;

[0010] S22. Construct a pre-denoising module: The pre-denoising module sequentially passes through an initial convolutional layer, a downsampling convolutional layer, a Transformer module, an upsampling and feature recovery layer, and an output refinement layer from input to output; the units in the downsampling convolutional layer and the units in the upsampling and feature recovery layer are connected by residual connections; the input of the pre-denoising module and the output refinement layer are connected by a residual connection;

[0011] S23. Construct a wavelet-Fourier enhancement module: The input F of the wavelet-Fourier enhancement module in is respectively input into the wavelet path and the Fourier path. The output of the wavelet path is merged into the Fourier path. The output of the Fourier path passes through weighted summation and then respectively passes through a phase component fusion stage and an amplitude component fusion stage. The outputs of the phase component fusion stage and the amplitude component fusion stage pass through an inverse Fourier transform module to obtain the output of the inverse Fourier transform module. The output of the inverse Fourier transform module is connected to F through a residual connection in to obtain the output F of the wavelet-Fourier enhancement module out ;

[0012] S24. Construct a feature fusion module: The feature fusion module sequentially passes through a connection layer, a 1×1 convolutional layer, a depthwise separable convolutional layer, and a 1×1 convolutional layer from input to output. I input and the output of the feature fusion module are connected by a residual connection;

[0013] S25. Define a loss function, and the loss function includes: a smooth L1 loss function L smooth and a perceptual loss function L per ;

[0014] S3. Input I onput into the image enhancement model for iterative training. When the number of iterative training reaches 200 times, the trained image enhancement model is obtained;

[0015] S31. Input I input into the YUV conversion module to obtain Y, U, and V channel images;

[0016] S32. Input the Y, U, and V channel images into three pre-denoising modules for denoising respectively to obtain the denoised channel images Y dn 、U dn and V dn ;

[0017] S33. Perform enhancement preprocessing on Y dn 、U dn and V dn respectively to obtain the input F of the wavelet-Fourier enhancement modulein ;

[0018] S34. Input F in into the wavelet-Fourier enhancement module to obtain the enhanced channel images Y enh , U enh and V enh , where {Y enh , U enh , V enh} ∈ F out ;

[0019] S35. Input the enhanced channel images Y enh , U enh and V enh into the feature fusion module to obtain the enhanced image I output ;

[0020] S4. Input the low-light image into the trained image enhancement model to obtain the enhanced image.

[0021] Further, the inverse image of the low-light image I input is 1 - I input ; The 1 - I input is extracted through a 3×3 convolutional layer and a ReLU function to obtain the feature map of the inverse image.

[0022] Further, the initial convolutional layer is specifically: a 3×3 convolutional layer combined with a ReLU activation function; the downsampling convolutional layer is specifically: three 3×3 convolutional layers combined with a ReLU activation function; the upsampling and feature recovery layer is specifically: three depthwise separable convolutional layers; the output refinement layer is specifically: a 3×3 convolutional layer combined with a Sigmoid activation function.

[0023] Further, in the wavelet-Fourier enhancement module, after F in inputs the wavelet path, it goes through two global pooling operations in sequence. The result of the first global pooling passes through the first Transformer module to obtain the fused feature of the first F in , and the result of the second global pooling passes through the second Transformer module to obtain the fused feature of the second F in ;

[0024] The first F inThe fused feature is input into the first wavelet fusion module. In the first wavelet fusion module: The input of the wavelet fusion module passes through a two-dimensional discrete wavelet transform module to obtain a low-frequency subband and a high-frequency subband. After connecting the low-frequency subband and the original small-scale feature, the output result obtained by successively passing through three 3×3 convolutional layers combined with the ReLU activation function and the output result obtained by passing the high-frequency subband through a 3×3 convolutional layer combined with the ReLU activation function are jointly input into the inverse wavelet transform module to obtain the output of the wavelet path; in the second wavelet fusion module, the original small-scale feature is the second F in 's fused feature.

[0025] Furthermore, in the wavelet-Fourier enhancement module, F in After being input into the Fourier path, it passes through the third Transformer module and the fast Fourier transform module respectively to obtain the fused feature of the third F in and the amplitude component A in and the phase component P in ; A in and P in pass through a 1×1 convolutional layer, a LeakyReLU activation function layer, and a 1×1 convolutional layer respectively to obtain the amplitude enhancement component A f and the phase enhancement component P f ;

[0026] The fused feature of the third F in is input into the second wavelet fusion module to obtain the output of the second wavelet fusion module; in the second wavelet fusion module, the original small-scale feature is the output of the wavelet path; the output of the second wavelet fusion module passes through the fast Fourier transform module again to obtain the amplitude wavelet fusion component A w and the phase wavelet fusion component P w ; A f and A w and P f and P w After weighted summation respectively, they enter the amplitude component fusion stage and the phase component fusion stage to obtain the amplitude fusion output A out and the phase fusion output P out .

[0027] Furthermore, the amplitude component fusion stage passes through three Transformer modules in sequence from input to output; the phase component fusion stage passes through four SCConv modules in sequence from input to output; there is a residual connection between the input and output of the amplitude component fusion stage; there is a residual connection between the input and output of the phase component fusion stage.

[0028] Furthermore, the F in ={Y′ dn ,U′dn , V' dn}, where Y' dn is obtained by connecting the feature map of the inverse image and Y dn , U' dn is obtained by connecting Y enh and U dn , and V' dn is obtained by connecting U enh and V dn .

[0029] Furthermore, the formula for calculating the smooth L1 loss function L smooth is: where smooth L1 () represents the smooth L1 loss function, x and y represent the center coordinates of the target bounding box, w and h respectively represent the width and height dimensions of the target bounding box, represents the value of the real image in the i-th dimension, represents the value of the enhanced image in the i-th dimension; the formula for calculating the perceptual loss function L per is: where C j , H j and W j respectively represent the number of channels, height, and width of the feature map output by the j-th convolutional layer in the convolutional neural network of the perceptual loss function Lper, φ j (I true ) represents the feature of the real image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper, and φ j (I output ) represents the feature of the enhanced image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper.

[0030] The beneficial effects of the present invention are as follows:

[0031] (1) By combining wavelet transform and Fourier transform in the YUV color space, the present invention makes full use of frequency domain information and achieves excellent performance on low-light image datasets. Compared with the prior art, it can better restore image details, adjust image saturation, reduce noise, and improve visual effects.

[0032] (2) The design of the pre-denoising module and the feature fusion module in the present invention further improves the image quality and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is the flowchart of the method described in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0034] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0035] This embodiment provides a low-light image enhancement method based on wavelet-Fourier transform. The flowchart of the method is as Figure 1 shown, and the method includes the following steps:

[0036] S1. Obtain an image dataset, which includes: a low-light image I input and a real image I true .

[0037] The inverse image of the low-light image I input is 1 - I input ; the 1 - I input is extracted through a 3×3 convolutional layer and a ReLU function (Conv(1)3×3 + ReLU) to obtain the feature map of the inverse image. The low-light image I input , the real image I true and the feature map of the inverse image can be used to train and verify the image enhancement model.

[0038] S2. Construct an image enhancement model with a multi-branch parallel architecture: The image enhancement model sequentially passes through a YUV conversion module, three pre-denoising modules, three wavelet-Fourier enhancement modules, and a feature fusion module from input to output; the parameters of each module in the image enhancement model are updated through a loss function;

[0039] S21. Construct a YUV conversion module: The YUV conversion formula is: Y = 0.999R + 0.587G + 0.114B, U = -0.14713R - 0.28886G + 0.436B, V = 0.615R + 0.51499G + 0.10001B, where Y represents luminance, U represents chrominance, V represents saturation, R represents red, G represents green, and B represents blue.

[0040] The YUV conversion module converts the input low-light image from the RGB color space to the YUV color space.

[0041] S22. Construct a pre-denoising module: The pre-denoising module sequentially passes through an initial convolutional layer, a downsampling convolutional layer, a Transformer module, an upsampling and feature recovery layer, and an output refinement layer from input to output; each unit in the downsampling convolutional layer and each unit in the upsampling and feature recovery layer are connected by a residual connection; the input of the pre-denoising module and the output refinement layer are connected by a residual connection.

[0042] The pre-denoising module is used for preliminary denoising. The initial convolutional layer is specifically: a 3×3 convolutional layer combined with a ReLU activation function (Conv(1) 3×3 + ReLU), with a stride of 1. The initial convolutional layer is used to extract preliminary features. The downsampling convolutional layer is specifically: three 3×3 convolutional layers combined with a ReLU activation function (Conv(2) 3×3 + ReLU), with a stride of 2. The downsampling convolutional layer is used to capture features at different scales. The Transformer module is at the network bottleneck, using a channel self-attention mechanism instead of the traditional spatial self-attention to reduce computational complexity and enhance the robustness of feature extraction. The upsampling and feature recovery layer is specifically: three depthwise separable convolutional layers (DWConv), with a stride of 2. The upsampling and feature recovery layer is used to restore the feature map to its original size. The output refinement layer is specifically: a 3×3 convolutional layer combined with a Sigmoid activation function (Conv(1) 3×3 + Sigmoid).

[0043] S23. Construct a wavelet-Fourier enhancement module: The input F of the wavelet-Fourier enhancement module in is respectively input into the wavelet path and the Fourier path. The output of the wavelet path is merged into the Fourier path. The output of the Fourier path, after weighted summation, passes through the phase component fusion stage and the amplitude component fusion stage respectively. The outputs of the phase component fusion stage and the amplitude component fusion stage then pass through an inverse Fourier transform module to obtain the output of the inverse Fourier transform module. The output of the inverse Fourier transform module is connected to F through a residual connection in to obtain the output F of the wavelet-Fourier enhancement module out .

[0044] In the wavelet-Fourier enhancement module, F in , after being input into the wavelet path, successively undergoes two global pooling (Globalpool) operations. The result of the first global pooling passes through the first Transformer module to obtain the fused feature of the first F in , and the result of the second global pooling passes through the second Transformer module to obtain the fused feature of the second F in . The Transformer module can handle long-range dependencies and reduce the loss of structural information during the downsampling process.

[0045] Take the first F inThe fused features are input into the first wavelet fusion module. In the first wavelet fusion module: the input of the wavelet fusion module passes through a two-dimensional discrete wavelet transform module (DWT) to obtain a low-frequency subband (LL) and high-frequency subbands (HH, HL, LH). After connecting the low-frequency subband and the original small-scale features, the output results obtained by successively passing through three 3×3 convolutional layers combined with the ReLU activation function and the output results obtained by passing the high-frequency subbands through a 3×3 convolutional layer combined with the ReLU activation function are jointly input into the inverse wavelet transform module (IDWT) to improve the accuracy and efficiency of feature fusion, and finally the output of the wavelet path is obtained; in the second wavelet fusion module, the original small-scale feature is the second F in fused features.

[0046] In the wavelet-Fourier enhancement module, F in After inputting the Fourier path, it passes through the third Transformer module and the fast Fourier transform module (FFT) respectively to obtain the third F in fused features and the amplitude component A in and the phase component P in ; A in and P in respectively pass through a 1×1 convolutional layer, a LeakyReLU activation function layer, and a 1×1 convolutional layer to obtain the amplitude enhancement component A f and the phase enhancement component P f .

[0047] Input the fused features of the third F in into the second wavelet fusion module to obtain the output of the second wavelet fusion module; in the second wavelet fusion module, the original small-scale feature is the output of the wavelet path; the output of the second wavelet fusion module passes through the fast Fourier transform module again to obtain the amplitude wavelet fusion component A w and the phase wavelet fusion component P w , define learnable parameters α and β to control the components that contribute less in the fusion process of the two-path outputs; A f and A w and P f and P w respectively pass through weighted summation and then enter the amplitude component fusion stage and the phase component fusion stage to obtain the amplitude fusion output A out and the phase fusion output P out .

[0048] The amplitude component fusion stage sequentially passes through three Transformer modules from input to output to facilitate capturing and integrating relevant information of the amplitude component; the phase component fusion stage sequentially passes through four SCConv modules from input to output to better understand the context information in the frequency domain; there is a residual connection between the input and output of the amplitude component fusion stage; there is a residual connection between the input and output of the phase component fusion stage.

[0049] The wavelet-Fourier enhancement module guides the amplitude information of the Fourier transform through the wavelet low-frequency information to achieve global illumination adjustment, and at the same time uses the Fourier phase information to precisely control the texture and details.

[0050] S24. Construct a feature fusion module: The feature fusion module sequentially passes through a connection layer, a 1×1 convolutional layer (Conv1×1), a depthwise separable convolutional layer, and a 1×1 convolutional layer from input to output. input There is a residual connection between I and the output of the feature fusion module; the depthwise separable convolutional layer effectively fuses the features of the three channels, reducing the computational amount. 1×1 convolutions are used before and after the depthwise separable convolutional layer for feature alignment and content fusion, and the content of I is implicitly enhanced and fused into the output image through the residual connection to enhance the global information, and finally the enhanced image I is obtained. input Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). output Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). output Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). enh Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). input Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). enh Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). enh Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). enh Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V). enh Specifically: I = Conv1×(DWConv(Conv1×1(I)))+I, where I = Concat(Y,U,V).

[0051] The feature fusion module integrates the information enhanced in the frequency domain and reconstructs the enhanced image. Through a large number of experiments, it is verified that the present invention exceeds the existing state-of-the-art low-light image enhancement methods on multiple data sets, showing excellent performance.

[0052] S25. Define a loss function for measuring the difference between the enhanced image and the real image. The loss function includes: a smooth L1 loss function L and a perceptual loss function L; smooth S25. Define a loss function for measuring the difference between the enhanced image and the real image. The loss function includes: a smooth L1 loss function L and a perceptual loss function L; per S25. Define a loss function for measuring the difference between the enhanced image and the real image. The loss function includes: a smooth L1 loss function L and a perceptual loss function L;

[0053] The calculation formula of the smooth L1 loss function L is: smooth The calculation formula of the smooth L1 loss function L is: where smooth() represents the smooth L1 loss function, L1 where smooth() represents the smooth L1 loss function, d represents I - I true d represents I - I output, x and y represent the center point coordinates of the target bounding box, and w and h represent the width dimension and height dimension of the target bounding box respectively. represents the value of the real image in the i-th dimension. represents the value of the enhanced image in the i-th dimension; the perceptual loss function L per is calculated as follows: where C j , H j and W j respectively represent the number of channels, height, and width of the output feature map of the j-th convolutional layer in the convolutional neural network (VGG19) of the perceptual loss function Lper, and φ j (I true ) represents the feature of the real image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper, and φ j (I output ) represents the feature of the enhanced image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper.

[0054] S3. Input I input into the image enhancement model for iterative training. When the number of iterative training reaches 200 times, the trained image enhancement model is obtained.

[0055] S31. Input I input into the YUV conversion module to obtain Y, U, and V channel images.

[0056] S32. Input the Y, U, and V channel images into three pre-denoising modules for denoising respectively to obtain the denoised channel images Y dn , U dn and V dn ;

[0057] S33. Perform enhancement preprocessing on Y dn , U dn and V dn respectively to obtain the input F in of the wavelet-Fourier enhancement module;

[0058] S34. Input F in into the wavelet-Fourier enhancement module to obtain the enhanced channel images Y enh , U enh and V enh , where {Y enh , U enh , V enh} ∈ F out ; the F in = {Y d ′ n , U′ dn , V d ′ n} where Y d ′ n is obtained by concatenating the feature map of the inverse image and Y dn U is obtained by concatenating ′ dn V is obtained by concatenating Y enh and U dn V is obtained by concatenating U d ′ n and V enh and V dn is obtained by concatenating.

[0059] S35. Input the enhanced channel images Y enh , U enh and V enh into the feature fusion module to obtain the enhanced image I output ;

[0060] S4. Input the low-light image into the trained image enhancement model to obtain the enhanced image.

Claims

1. A low-light image enhancement method based on wavelet-Fourier transform, characterized in that, The method includes the following steps: S1. Obtain an image dataset, where the image dataset includes: a low-light image I input and a ground-truth image I true ; S2. Construct an image enhancement model: The image enhancement model sequentially passes through a YUV conversion module, three pre-denoising modules, three wavelet-Fourier enhancement modules, and a feature fusion module from input to output; the parameters of each module in the image enhancement model are updated through a loss function. S21. Construct a YUV conversion module: The YUV conversion formula is: Y = 0.999R + 0.587G + 0.114B, U = -0.14713R - 0.28886G + 0.436B, V = 0.615R + 0.51499G + 0.10001B, where Y represents brightness, U represents chrominance, V represents saturation, R represents red, G represents green, and B represents blue. The YUV conversion module converts the input low-light image from the RGB color space to the YUV color space. S22. Construct a pre-denoising module: The pre-denoising module sequentially passes through an initial convolutional layer, a downsampling convolutional layer, a Transformer module, an upsampling and feature recovery layer, and an output refinement layer from input to output; each unit in the downsampling convolutional layer and each unit in the upsampling and feature recovery layer are connected through residual connections; the input of the pre-denoising module and the output refinement layer are connected through a residual connection. S23. Construct a wavelet-Fourier enhancement module: The input F of the wavelet-Fourier enhancement module in is respectively input into the wavelet path and the Fourier path. The output of the wavelet path is merged into the Fourier path. The output of the Fourier path is weighted and summed and then passes through the phase component fusion stage and the amplitude component fusion stage respectively. The outputs of the phase component fusion stage and the amplitude component fusion stage pass through the inverse Fourier transform module again to obtain the output of the inverse Fourier transform module. The output of the inverse Fourier transform module is connected to F through a residual connection in to obtain the output of the wavelet-Fourier enhancement module out ; S24. Construct a feature fusion module: The feature fusion module sequentially passes through a connection layer, a 1×1 convolutional layer, a depthwise separable convolutional layer, and a 1×1 convolutional layer from input to output. A residual connection is made between I input and the output of the feature fusion module; S25. Define a loss function, where the loss function includes: a smooth L1 loss function L smooth and a perceptual loss function L per ; S3. Input I input into the image enhancement model for iterative training. When the number of iterative training reaches 200 times, the trained image enhancement model is obtained; S31. Input I input into the YUV conversion module to obtain Y-channel, U-channel, and V-channel images; S32. Input the Y, U, and V channel images into three pre-denoising modules respectively for denoising to obtain the denoised channel images Y dn , U dn and V dn ; S33. Enhance and preprocess Y dn , U dn and V dn respectively to obtain the input F in ; S34. Input F in into the wavelet-Fourier enhancement module to obtain the enhanced channel images Y enh , U enh and V enh , where {Y enh , U enh , V enh} ∈ F out ; S35. Input the enhanced channel images Y enh , U enh and V enh into the feature fusion module to obtain the enhanced image I output ; S4. Input the low-light image into the trained image enhancement model to obtain the enhanced image.

2. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, characterized in that, The low-light image I input has an inverse image of 1 - I input ; the 1 - I input is extracted through a 3×3 convolutional layer and a ReLU function to obtain the feature map of the inverse image.

3. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, characterized in that, The initial convolutional layer is specifically: a 3×3 convolutional layer combined with a ReLU activation function; the downsampling convolutional layer is specifically: three 3×3 convolutional layers combined with a ReLU activation function; the upsampling and feature recovery layer is specifically: three depthwise separable convolutional layers; the output refinement layer is specifically: a 3×3 convolutional layer combined with a Sigmoid activation function.

4. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, wherein In the wavelet-Fourier enhancement module, F in After inputting the wavelet path, it goes through two global pooling operations in sequence. The result of the first global pooling passes through the first Transformer module to obtain the fused feature of the first F tn The result of the second global pooling passes through the second Transformer module to obtain the fused feature of the second F in ; Input the fusion feature of the first F in into the first wavelet fusion module. In the first wavelet fusion module: the input of the wavelet fusion module undergoes a two-dimensional discrete wavelet transform module to obtain a low-frequency subband and a high-frequency subband. After connecting the low-frequency subband and the original small-scale feature, the output result obtained by successively passing through three 3×3 convolutional layers combined with the ReLU activation function and the output result obtained by passing the high-frequency subband through a 3×3 convolutional layer combined with the ReLU activation function are jointly input into the inverse wavelet transform module to obtain the output of the wavelet path; in the second wavelet fusion module, the original small-scale feature is the fusion feature of the second F in ​ 5. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, characterized in that In the wavelet-Fourier enhancement module, F in After inputting the Fourier path, it passes through the third Transformer module and the fast Fourier transform module respectively to obtain the fused feature and amplitude component A in of the third F in and phase component P in ; A in and P in respectively pass through a 1×1 convolutional layer, a LeakyReLU activation function layer, and a 1×1 convolutional layer in sequence to obtain the amplitude-enhanced component A f and the phase-enhanced component P f ; Input the fusion feature of the third F in into the second wavelet fusion module to obtain the output of the second wavelet fusion module; in the second wavelet fusion module, the original small-scale feature is the output of the wavelet path; the output of the second wavelet fusion module then passes through a fast Fourier transform module to obtain the amplitude wavelet fusion component A w and the phase wavelet fusion component P w ; A f and A w and P f and P w respectively pass through weighted summation and then enter the amplitude component fusion stage and the phase component fusion stage to obtain the amplitude fusion output A out and the phase fusion output P out .

6. The low-light image enhancement method based on wavelet-Fourier transform according to claim 5, characterized in that The amplitude component fusion stage sequentially passes through three Transformer modules from input to output; the phase component fusion stage sequentially passes through four SCConv modules from input to output; the input and output of the amplitude component fusion stage are connected through a residual connection; the input and output of the phase component fusion stage are connected through a residual connection.

7. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, characterized in that The said F in = {Y′ dn , U′ dn , V′ dn}, where Y′ dn is obtained by connecting the feature map of the inverse image and Y dn , U′ dn is obtained by connecting Y enh and U dn , and V′ dn is obtained by connecting U enh and V dn .

8. The low-light image enhancement method based on wavelet-Fourier transform according to claim 1, wherein The smooth L1 loss function L smooth is calculated as follows: Among them, smooth L1 () represents the smooth L1 loss function, x and y represent the center point coordinates of the target bounding box, and w and h respectively represent the width dimension and height dimension of the target bounding box, represents the value of the real image in the i-th dimension, represents the value of the enhanced image in the i-th dimension; the perceptual loss function L per The calculation formula of is: Among them, C j , H j and W j respectively represent the number of channels, height, and width of the output feature map of the j-th convolutional layer in the convolutional neural network of the perceptual loss function Lper, φ j (I trie ) represents the feature of the real image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper, φ j (I output ) represents the feature of the enhanced image extracted by the j-th convolutional layer of the convolutional neural network of the perceptual loss function Lper.

Images