Low-illumination image enhancement method based on YCbCr space and Fourier frequency domain

Abstract

The invention discloses a low-illumination image enhancement method based on a YCbCr space and a Fourier frequency domain, and the method comprises the steps: extracting a brightness component of a low-illumination image in the YCbCr space, and constructing a training set together with a normal-illumination image and the low-illumination image; constructing a low-illumination image enhancement network model, and inputting the training set into the model for training; constructing a loss function between the amplitude enhancement network branch and the normal illumination image as well as between the phase enhancement network branch and the normal illumination image, and adjusting model parameters by minimizing the loss function to obtain an optimization model; and inputting a to-be-enhanced low-illumination image into the optimization model for image enhancement. According to the method, the YCbCr brightness component and the amplitude information of the Fourier frequency domain are combined, and the Transform is utilized to enhance the phase information, so that the visual quality of the low-illumination image is remarkably improved, the visibility and definition of the image are improved, and the method has wide application prospects in tasks of image enhancement, denoising, super-resolution recovery and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of low-light image enhancement methods, and particularly relates to a low-light image enhancement method based on YCbCr space and Fourier frequency domain. Background Art

[0002] In real-world environments, lighting conditions are often less than ideal. Due to environmental factors and technical limitations, low-light conditions are common. This leads to degraded image quality and loss of detail, negatively impacting human visual perception and subsequent visual tasks. Low-light image enhancement aims to improve lighting conditions to produce images that are closer to natural, normal-light images. By improving image quality in low-light environments, human visual perception can be effectively enhanced, resulting in better performance in a variety of visual tasks, such as object detection, image segmentation, and facial recognition.

[0003] Low-light image enhancement methods can be categorized into traditional methods and deep learning-based methods. Traditional methods can generally be divided into histogram equalization, gamma correction, and methods based on Retinex theory. Histogram equalization increases contrast by flattening the pixel value distribution. However, using the entire histogram information can lead to overexposure in certain parts of the image, affecting visual quality and potentially introducing artifacts. Gamma correction is a classic nonlinear image enhancement method. In this process, each pixel in the image is processed individually, and the same transformation operation is applied to all pixel values. However, this method often results in image color distortion and overexposure, and its performance is suboptimal in various scenarios. Methods based on Retinex theory provide an intuitive physical basis for low-light image enhancement and are more consistent with the human eye's visual mechanism. Consequently, many related algorithms have been derived, but they lack the ability to improve image color and texture detail.

[0004] Deep learning-based methods can automatically extract and learn complex features and patterns in low-light images from large datasets by building multi-layer neural network models and training them with large amounts of data, effectively addressing image degradation. LLNet, proposed by Lore et al., is the first deep learning-based low-light image enhancement method. It uses a convolutional neural network (CNN) to achieve end-to-end contrast enhancement and denoising of images. PPformer, proposed by Dang et al., is a lightweight image enhancement network. This network effectively models the dependencies between local and non-local regions by calculating a dual cross-attention mechanism. Using this approach, the network is able to restore images and improve their color consistency, natural brightness, and contrast.

[0005] Several studies have shown that there is a positive correlation between the amplitude component of the Fourier transform and image brightness, while the phase component represents the structural information of the image and is not easily affected by brightness changes. In addition, the Fourier transform can extract global information from the image without introducing a large number of neural network parameters like multi-layer perceptrons or transformers. Therefore, using the Fourier transform to perform image brightness enhancement in the frequency domain is an effective option. In addition, the YCbCr space is also considered a good choice for performing texture-color decomposition, and the brightness component (Y) of the YCbCr color space performs well and is least affected by noise. Therefore, a method combining the YCbCr color space and the Fourier transform is used to enhance low-light images. Summary of the Invention

[0006] The purpose of the present invention is to provide a low-light image enhancement method based on YCbCr space and Fourier frequency domain, which solves the problem that the existing image enhancement methods have poor effect on improving image quality in low-light environments.

[0007] The technical solution adopted by the present invention is: a low-light image enhancement method based on YCbCr space and Fourier frequency domain, comprising the following steps:

[0008] Step 1: Extract the brightness component of the low-light image in the YCbCr space and construct a training set together with the normal-light image and the low-light image;

[0009] Step 2: Build a low-light image enhancement network model including an amplitude enhancement network branch, a phase enhancement network branch, and a feature fusion module, and input the training set obtained in step 1 into the model for training;

[0010] Step 3: Based on the mean square error loss and perceptual loss, a loss function is constructed between the amplitude enhancement network branch and the phase enhancement network branch in the model obtained in step 2 and the normal light image. The model parameters are adjusted by minimizing the loss function to obtain an optimized low-light image enhancement network model.

[0011] Step 4: Input the low-light image to be enhanced into the model obtained in step 3 for image enhancement.

[0012] The present invention is also characterized in that:

[0013] Step 2 specifically includes the following steps:

[0014] Step 2.1: The low-light image I in the training set low and paired brightness component Y low The amplitude enhancement network branch is input together to obtain the enhanced image I after brightness improvement. out1 ;

[0015] Step 2.2: The low-light image I in the training set low The input is fed into two serial phase enhancement network branches. The output of the first branch is fed into the input of the second branch, and the enhanced image I with restored detail structure information is obtained. out2 ;

[0016] Step 2.3: Image I obtained in step 2.1 out1 and image I obtained in step 2.2 out2 Input feature fusion module, adaptively fuse the different features of the two branches, and obtain the enhanced image I out .

[0017] Step 2.1 specifically includes the following steps:

[0018] Step 2.1.1: Convert the brightness component Y of the low-light image low Input the brightness prediction network LumiNet to predict the Y component of the normal lighting image in the YCbCr space and obtain the predicted component Y′:

[0019] Y′=LumiNet(Y low )

[0020] The predicted component Y′ is then converted to the Fourier domain through Fourier transform, and the amplitude component Y in the Fourier domain is extracted. lum , the Fourier transform formula is:

[0021]

[0022] Among them, X(u,v) is the complex component, H and W are the height and width of the input image, h and w are the coordinates in the spatial domain, u and v are the coordinates in Fourier space, j is the imaginary unit, and the inverse Fourier transform process is represented by F -1 ;

[0023] Each complex component X(u,v) is represented by the amplitude component A(X(u,v)) and the phase component P(X(u,v)) as follows:

[0024]

[0025] Where R(X) and I(X) represent the real and imaginary parts of X(u,v), respectively;

[0026] For the amplitude component Y lum After two 1x1 convolutions, LeakyReLU activation function and SELayer feature extraction, the feature components are obtained The specific process formula is as follows:

[0027] Y′ lum =Conv(Y lum)

[0028] Y" lum =LeakyReLU(Y′ lum )

[0029] Y″′ lum =SELayer(Y″) lum )

[0030]

[0031] Among them, Y′ lum , Y″ lum and Y″′ lum is the extracted intermediate feature, is the characteristic component of the final output;

[0032] Step 2.1.2: Low-light image I low and characteristic components The input is sent to the amplitude enhancement module AmpEnhancer for amplitude enhancement in the Fourier domain. The amplitude enhancement module AmpEnhancer contains six enhancement blocks. Each enhancement block outputs a further enhanced image as the input of the next enhancement block. Jump connections are made between the two enhancement blocks. For the low-light image I input to the enhancement block low First, perform Fourier transform to the frequency domain and extract the amplitude component A low and phase component P low , and then component A low and P Low Feature extraction is performed through two 1x1 convolutions, LeakyReLU activation function and SELayer to obtain A′ low and P′ low ;

[0033] Step 2.1.3: Component A′ low Use the sigmoid activation function to limit A′ to the range of (0,1). low and Y′ lum After pixel-by-pixel addition and Sigmoid(A′ low ) pixel by pixel, and then perform residual connection to obtain the enhanced amplitude A that combines the feature extraction amplitude component of the low-light image and the feature extraction amplitude component of the YCbCr space out , then A out and P′ low After inverse Fourier transform, the output I′ of the enhancement block is obtained low :

[0034]

[0035] I′low =F -1 (A out , P′ low )

[0036] in, represents pixel-by-pixel addition, Indicates pixel-by-pixel multiplication, F -1 represents the inverse Fourier transform;

[0037] Finally, the output I′ of the current enhancement block is low Input to the next enhancement block and perform the same operation until the six enhancement blocks are completed, and the enhanced image I with improved brightness is obtained. out1 .

[0038] Step 2.2 specifically includes the following steps:

[0039] Step 2.2.1: First, the low-light image I low (b,c,h,w) is reshaped, and the image dimension is transformed into (b,h*w,c). Then the transformed image is input into the layer normalization module to normalize the channel number dimension. The processed tensor is then restored to its shape to obtain I ln (b,c,h,w);

[0040] Step 2.2.2, normalize the feature I ln (b,c,h,w) is input into the phase-guided attention module PGA to obtain the attention-weighted feature I attn (b,c,h,w); specifically: I ln (b,c,h,w) first passes through a convolution operator with a kernel size of 1x1 to generate a feature map, and the number of channels is expanded by 3 times to obtain (b,3*c,h,w). Then, it passes through a 3x3 deep convolution layer to further process the features. Then, it is evenly split into three parts along the dimension of the number of channels, which are used as the query (q(b,c,h,w)), key (k(b,c,h,w)) and value (v(b,c,h,w)) of the attention module respectively:

[0041] k,q,v=Conv3(Conv1(I in ))

[0042] Among them, Conv1 represents 1x1 convolution, and Conv3 represents 3x3 convolution;

[0043] Step 2.2.3: Input the obtained q and k components into the phase enhancement module PB to extract the detailed outline contained in the phase; specifically: in the phase enhancement module, the input q(b,c,h,w) and k(b,c,h,w) are first subjected to a 1x1 convolution to perform preliminary feature adjustment, and then the phase components P of q and k are extracted through Fourier transform. q and P k , and then the phase component P q and P k Input into two 1x1 convolutions, LeakyReLU activation function and SELayer layer to get the enhanced phase P′ q and P′ k , and then the enhanced phase and the amplitude set to 1 are transformed from the frequency domain to the spatial domain through inverse Fourier transform to obtain φ q and φ k :

[0044] P q =FFT(Conv(q)), P k =FFT(Conv(k))

[0045] P′ q =Conv(SE(Lea(Conv(P q )))),P′ k =Conv(SE(Lea(Conv(P k ))))

[0046] φ q =iFFT(P′ q ),φ k =iFFT(P′ k )

[0047] Among them, Conv represents 1x1 convolution, SE represents SELayer, Lea represents LeakyReLU activation function, Y q and Y k Represents the generated intermediate variables;

[0048] Step 2.2.4: Output φ of the phase enhancement module q (b,c,h,w) and φ k (b,c,h,w), and v(b,c,h,w) are reshaped and L2 normalized to become and Make the L2 norm of each vector 1, and then and Input to the attention module for context enrichment, specifically: first calculate and The similarity between them is measured by matrix multiplication to obtain an attention weight matrix of shape (b, 1, h*w, h*w), which represents the similarity between each pixel. Then the similarity matrix is passed through the softmax function and multiplied by the scaling factor temperature to scale the attention weight. Perform weighted summation and reshape to (b, c, h, w) to obtain the final output I of the phase-guided attention module attn (b,c,h,w):

[0049]

[0050] Among them, PB represents the phase enhancement module, R represents the dimension reshaping process, represents matrix multiplication, and temp represents the scaling factor;

[0051] Step 2.2.5: Direct the output of the attention module I attn (b,c,h,w),I low (b,c,h,w) and I ln (b,c,h,w) are added to get the output Y(b,c,h,w):

[0052]

[0053] Then input Y(b,c,h,w) into the layer normalization module to obtain Y ln , and then Y ln Input into the feedforward neural network FRFN to get the output Y fn :

[0054] Y ln =LN(Y)

[0055] T fn =FRFN(Y ln )

[0056] Finally, Y, Y ln and Y fn Perform element-by-element addition to obtain the output I of the second phase enhancement network branch out2 :

[0057]

[0058] Step 2.3 is as follows: the image I obtained in step 2.1 is out1 and image I obtained in step 2.2 out2 Input the feature fusion module SKFusion to obtain the enhanced image I out :

[0059] Iout =SK(I out1 , I out2 ).

[0060] The loss function constructed in step 3 is expressed as:

[0061] L total =L s1 +λ1L vgg +λ2L s2 +λ3L s3 +λ4L lum

[0062] Among them, λ1, λ2, λ3 and λ4 represent loss weights; L s1 Represents the enhanced image I output by the low-light image enhancement network model out And the L2 loss between the real normal lighting image GT:

[0063] L s1 =||I out -GT||2

[0064] L vgg Represents the enhanced image I output by the low-light image enhancement network model out The perceptual loss between the real normal lighting image GT constrains the features extracted from VGG; L s2 represents the output of the amplitude enhancement network branch I out1 And the L2 loss between the real normal lighting image GT:

[0065] L s2 =||I out1 -GT||2

[0066] L s3 I represents the output of the phase enhancement network branch out2 And the L2 loss between the real normal lighting image GT:

[0067] L s3 =||I out2 -GT||2

[0068] L lum Represents the output of the brightness prediction network LumiNet and the Y component Y of the normal lighting image YCbCr space high The L2 loss between:

[0069] L lum =||LumiNet(Y low )-Y high ||2.

[0070] The loss weights of the loss function are set to λ1,λ2,λ3,λ4 = [0.3, 0.2, 0.2, 0.1].

[0071] Step 4 specifically includes the following steps:

[0072] Step 4.1: First, convert the low-light image to be enhanced into YCbCr space to extract the brightness component Y low , and then input it into the amplitude enhancement network branch together with the low-light image to obtain a brightness-enhanced image;

[0073] Step 4.2: Input the low-light image to be enhanced into the phase enhancement network branch to restore the detail structure information and obtain an image with restored texture details;

[0074] Step 4.3: Feature fusion is performed on the images output by the amplitude enhancement network branch and the phase enhancement network branch to obtain an enhanced image.

[0075] The beneficial effects of the present invention are as follows: the low-light image enhancement method based on the YCbCr space and Fourier frequency domain of the present invention can significantly improve the brightness table of the image by separating the brightness component in the YCbCr color space and converting it into the Fourier domain for processing, and combining the amplitude information of the low-light image in the Fourier domain. In this process, the recovery of the brightness component effectively improves the problem of insufficient image brightness in low-light environments; in addition, the present invention also extracts the phase component in the low-light image through Fourier transform, and uses the Transformer architecture to enhance the phase information. The self-attention mechanism of the Transformer can optimize the phase information of the image globally, and restore the detailed structure of the image by enhancing the phase, especially in terms of the edges and textures of the image, which can effectively improve the detail performance of the image.

[0076] In summary, the present invention significantly improves the visual quality of low-light images and the visibility and clarity of images by combining the YCbCr luminance component and the amplitude information of the Fourier frequency domain, and using the Transformer to enhance the phase information. It has broad application prospects in tasks such as image enhancement, denoising, and super-resolution restoration. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 1 is a flow chart of a low-light image enhancement method based on YCbCr space and Fourier frequency domain according to the present invention;

[0078] Figure 2 This is a schematic diagram of the overall dual-branch structure of the present invention based on YCbCr space and Fourier frequency domain;

[0079] Figure 3Schematic diagram of the network structure of the amplitude enhancement branch combining YCbCr space and Fourier frequency domain of the present invention;

[0080] Figure 4 Schematic diagram of the network structure of the phase enhancement branch combining Fourier domain and Transformer in the present invention;

[0081] Figure 5 This is a network structure diagram of the amplitude enhancement module (AmpEnhancer) of the present invention;

[0082] Figure 6 1 is a network structure diagram of the phase-guided attention module (PGA) of the present invention;

[0083] Figure 7 Schematic diagram of the structure of the phase enhancement module (PB) in the phase-guided attention module of the present invention;

[0084] Figure 8 This is a diagram of the visual enhancement effect of the present invention on the real dataset of the benchmark dataset LOLv2. DETAILED DESCRIPTION

[0085] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0086] Example 1

[0087] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain, which is specifically implemented according to the following steps:

[0088] Step 1: For a public benchmark dataset containing paired normal-light images and low-light images, extract the luminance component (Y component) of the YCbCr space from the low-light image and construct a training set together with the normal-light image and the low-light image.

[0089] Step 2: Design a low-light image enhancement network. The low-light image enhancement network architecture includes a brightness enhancement module, a spatial detail recovery module, and a feature fusion module. The constructed training set is input into the image enhancement network, and the final output is an enhanced image with improved brightness and rich details. The specific process is as follows:

[0090] Step 2.1: The low-light image I in the training set low and paired brightness component Y low The two are input to the brightness enhancement module branch to obtain the enhanced image I after brightness improvement. out1 ;

[0091] Step 2.2: The low-light image I in the training set lowThe input is fed into two serial spatial detail recovery module branches. The output of the first module is the input of the second module, and the enhanced image I with image structure recovery is finally obtained. out2 ;

[0092] Step 2.3: The image with improved brightness I out1 Image I restored with detailed structural information out2 Input to the feature fusion module, adaptively fuse the different features of the two branches, and obtain the final enhanced image I out .

[0093] Step 3: Construct a loss function for the enhanced image and the normal illumination image based on the mean square error loss and perceptual loss. Improve the visual quality of the enhanced image by minimizing the pixel-level differences and deep feature differences between the enhanced image and the normal illumination image.

[0094] Step 4: Enhance the low-light image based on the optimized low-light image enhancement network.

[0095] Example 2

[0096] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain. Based on Example 1, the specific process of step 2.1 is preferably:

[0097] Step 2.1.1: Convert the Y low The component is input into the brightness prediction network (LumiNet), which predicts the Y component of the YCbCr space of the normal light image and obtains the predicted component Y'. The formula for this process is:

[0098] Y′=LumiNet(Y low )

[0099] Then the predicted component Y′ is transformed into the Fourier domain through Fourier transform, and the amplitude component of the component in the Fourier domain is extracted to obtain Y lum , the Fourier transform formula is:

[0100]

[0101] Where X(u,v) is the complex component, H and W are the height and width of the input image, h and w are the coordinates in the spatial domain, u and v are the coordinates in Fourier space, and j is the imaginary unit. The inverse Fourier transform process is represented by F -1 ;

[0102] Each complex component X(u,v) can be represented by an amplitude component A(X(u,v)) and a phase component P(X(u,v)). These two components are expressed as follows:

[0103]

[0104] Here, R(X) and I(X) represent the real and imaginary parts of X(u,v), respectively.

[0105] For the amplitude component Y lum After two 1x1 convolutions, LeakyReLU activation function and SELayer feature extraction, the feature components are obtained The specific process formula is as follows:

[0106] Y' lum =Conv(Y lum )

[0107] Y" lum =Leaky ReLU(Y' lum )

[0108] Y"′ lum =SELayer(Y" lum )

[0109]

[0110] where Y′ lum , Y″ lum and Y″′ lum is the extracted intermediate feature, It is the characteristic component of the final output.

[0111] Step 2.1.2: Low-light image I low and characteristic components Input to the amplitude enhancement module (AmpEnhancer) for amplitude enhancement in the Fourier domain. This module contains six enhancement blocks, each of which outputs a further enhanced image as the input of the next module, and jump connections are made between each two enhancement modules. For the low-light image I input to the enhancement module low First, perform Fourier transform to the frequency domain and extract the amplitude component A low and phase component P low , and then the two components are respectively subjected to two 1x1 convolutions, LeakyReLU activation function and SELayer for feature extraction to obtain A′ low and P′ low .

[0112] Step 2.1.3, A′ low Use the sigmoid activation function to limit it to the range of (0,1), and then A′ low and Y′ lum After pixel-by-pixel addition and Sigmoid(A′ low) pixel by pixel, and then perform residual connection to obtain the enhanced amplitude A that combines the feature extraction amplitude component of the low-light image and the feature extraction amplitude component of the YCbCr space out . Then A out and P′ low After inverse Fourier transform, the output I′ of the enhancement module is obtained low , the output will be input to the next enhancement module to perform the same operation as step 2.1.3 until the six enhancement modules are completed. The formula of this process is as follows:

[0113]

[0114] I′ low =F -1 (A out , P′ low )

[0115] in, represents pixel-by-pixel addition, Indicates pixel-by-pixel multiplication, F -1 represents the inverse Fourier transform.

[0116] Example 3

[0117] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain. Based on Example 1, the specific process of step 2.2 is preferably:

[0118] Step 2.2.1. Low-light image I low (b,c,h,w) is first reshaped to (b,h*w,c) so that the normalization operation can be performed independently on all pixel positions of each channel. The transformed image is then input into the layer normalization module to normalize the channel number dimension. The processed tensor is then reshaped to obtain I ln (b,c,h,w).

[0119] Step 2.2.2, normalize the feature I ln (b,c,h,w) is input into the phase-guided attention module (PGA) to obtain the attention-weighted feature I attn (b,c,h,w). Specifically, I lnFirst, a convolution operator with a kernel size of 1x1 is used to generate a feature map, and the number of channels is expanded to three times the original number (b, 3*c, h, w). Then, a 3x3 deep convolution layer is used to further process the features, and then it is evenly divided into three parts along the dimension of the number of channels, which serve as the query (q(b,c,h,w)), key (k(b,c,h,w)) and value (v(b,c,h,w)) of the attention module respectively.

[0120] k,q,v=Conv3(Conv1(I in )) Among them, Conv1 represents 1x1 convolution and Conv3 represents 3x3 convolution.

[0121] Step 2.2.3: In order to further enhance the structural features of q and k and enhance the phase of the image to obtain the detailed features of the image, the obtained q and k components are respectively input into the phase enhancement module (PB) to extract the detailed contours contained in the phase. Specifically, in the phase enhancement module, the input q(b,c,h,w) and k(b,c,h,w) are first subjected to a 1x1 convolution to perform preliminary adjustment on the features, and then subjected to Fourier transform to extract the phase components P of q and k. q and P k In order to further enhance the phase component, the phase component is input into two 1x1 convolutions, LeakyReLU activation function and SELayer layer to obtain the enhanced phase P′ q and P′ k , and the amplitude is set to 1. Then the enhanced phase and the amplitude set to 1 are transformed from the frequency domain to the spatial domain through inverse Fourier transform to obtain φ q and φ k To proceed to the next step. The formula of this process is as follows:

[0122] P q =FFT(Conv(q)), P k =FFT(Conv(k))

[0123] P′ q =Conv(SE(Lea(Conv(P q )))),P′ k =Conv(SE(Lea(Conv(P k ))))

[0124] φ q =iFFT(P′ q ),φ k =iFFT(P′ k )

[0125] Among them, Conv represents 1x1 convolution, SE represents SELayer, Lea represents LeakyReLU activation function, Y q and Y k Represents the generated intermediate variable.

[0126] Step 2.2.4: For the output φ of the phase enhancement module q (b,c,h,w) and φ k (b,c,h,w), and v(b,c,h,w) are reshaped and L2 normalized to become and Make the L2 norm of each of their vectors 1 to ensure that the influence of each position is on the same scale when calculating the attention weight. and Input to the attention module for context enrichment. The specific process is: first calculate and The similarity between them is calculated by matrix multiplication to obtain an attention weight matrix of shape (b,1,h*w,h*w), which represents the similarity between each pixel. Then the similarity matrix is passed through the softmax function and multiplied by the scaling factor temperature to scale the attention weight. Perform weighted summation and reshape to (b, c, h, w) to obtain the final output I of the phase-guided attention module attn (b,c,h,w). The formula for this step is as follows:

[0127]

[0128] Among them, PB represents the phase enhancement module, R represents the dimension reshaping process, represents matrix multiplication, and temp represents the scaling factor.

[0129] Step 2.2.5: Direct the output of the attention module I attn (b,c,h,w) and I low (b,c,h,w) and I ln (b,c,h,w) are added to obtain the further output Y(b,c,h,w), the formula is as follows:

[0130]

[0131] Then input Y(b,c,h,w) into the layer normalization module (Layer Normalization) to obtain Y ln , and then Y ln Input into the feedforward neural network FRFN to eliminate redundant features and improve the network's recovery ability to obtain the output Yfn :

[0132] Y ln =LN(Y)

[0133] Y fn =FRFN(Y ln )

[0134] Finally, Y, Y ln and Y fn Perform element-by-element addition to obtain the output I of the second phase enhancement branch out2 :

[0135]

[0136] Example 4

[0137] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain. Based on Example 1, the specific process of step 2.3 is preferably:

[0138] The output result of the first branch amplitude enhancement I out1 and the second branch phase enhancement result I out2 Input into the feature fusion module SKFusion to obtain the final low-light image enhancement result I out :

[0139] I out =SK(I out1 , I out2 )

[0140] Example 5

[0141] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain. Based on Example 1, the specific process of step 3 is preferably:

[0142] Based on the mean square error loss and perceptual loss, the loss function between the amplitude enhancement branch and the brightness enhancement branch and the normal illumination image is constructed, which consists of the following five parts:

[0143] L total =L s1 +λ1L vgg +λ2L s2 +λ3L s3 +λ4L lum

[0144] Among them, λ1, λ2, λ3 and λ4 represent loss weights, which are empirically set to λ1, λ2, λ3, λ4 = [0.3, 0.2, 0.2, 0.1].

[0145] L s1Indicates the final output I out The L2 loss between GT and real normal lighting images:

[0146] L s1 =||I out -GT||2

[0147] L vgg Indicates the final output I out The perceptual loss between L and GT constrains the features extracted from VGG. s2 Represents the output result of the amplitude enhancement branch I out1 L2 loss between GT and CNN:

[0148] L s2 =||I out1 -GT||2

[0149] L s3 Represents the output result of the phase enhancement branch I out2 L2 loss between GT and CNN:

[0150] L s3 =||I out2 -GT||2

[0151] L lum Represents the L2 loss between the output of the brightness prediction network (LumiNet) and the Y component of the YCbCr space of the normal light image:

[0152] L lum =||LumiNet(Y low )-Y high ||2

[0153] Among them, Y high Represents the Y component of the YCbCr space of the normal light image.

[0154] Example 6

[0155] The present invention provides a low-light image enhancement method based on YCbCr space and Fourier frequency domain, such as Figure 1 As shown, please follow the steps below:

[0156] Step 1: Use a public benchmark dataset containing paired normal light images and low light images to extract the brightness component Y in the YCbCr color space from the low light image. low , and reconstruct the training set together with normal light images and low light images. Each pair of images in the public benchmark dataset comes from the same scene and is used to train the low light image enhancement model.

[0157] Step 2: Build a low-light image enhancement network. The overall structure is as follows: Figure 2 As shown, the reconstructed training set is input into the low-light image enhancement network model for training to obtain an enhanced image; the low-light image enhancement network model includes an amplitude enhancement branch, a phase enhancement branch, and a feature fusion module. The specific process is:

[0158] Step 2.1: Compare the low-light image in the training set and the brightness component Y in the YCbCr color space low Input into the first branch amplitude enhancement branch to obtain the brightness enhanced image I out1 .

[0159] According to existing research, in Fourier space, brightness degradation is mainly reflected in the amplitude component. Therefore, by estimating the amplitude transform map in Fourier space, the brightness of low-light images can be effectively improved. In addition, among various color spaces, the brightness component (Y) of the YCbCr color space performs best and is the least sensitive to noise. Compared with other color spaces, the brightness component of YCbCr can more effectively preserve image details and reduce the impact of noise. Based on this, in the first branch, the brightness component of the YCbCr color space is combined with the amplitude component of the low-light image, which can significantly improve the brightness performance of the image without introducing additional noise. The specific process is:

[0160] Step 2.1.1, as Figure 3 As shown, the brightness component Y of the low-light image in the training set is low The input is fed into the luminance prediction network (LumiNet), a Unet network consisting of an encoder and a decoder. This network is able to preserve image detail and can be used to predict the Y component of the YCbCr space of the normal light image, yielding the predicted component Y′. The predicted luminance component contains more information close to the image brightness under normal light conditions, which helps restore the brightness of low-light images. The formula for this process is:

[0161] y′=LumiNet(Y low )

[0162] Then the predicted component Y′ is transformed into the Fourier domain through Fourier transform, and the amplitude component of the component in the Fourier domain is extracted to obtain Y lum , the Fourier transform formula is:

[0163]

[0164] Where X(u,v) is the complex component, H and W are the height and width of the input image, h and w are the coordinates in the spatial domain, u and v are the coordinates in Fourier space, and j is the imaginary unit. The inverse Fourier transform process is represented by F -1;

[0165] Each complex component X(u,v) can be represented by an amplitude component A(X(u,v)) and a phase component P(X(u,v)). These two components are expressed as follows:

[0166]

[0167] Here, R(X) and I(X) represent the real and imaginary parts of X(u,v), respectively.

[0168] For the amplitude component Y lum Feature extraction is performed through two 1x1 convolutions, LeakyReLU activation function and SELayer. Among them, the two convolution layers can extract local features in the amplitude image while maintaining the number of channels and dimensional information of the image. The LeakyReLU activation function performs a nonlinear transformation on the input to enhance the network's expressive power and prevent the gradient vanishing problem. SELayer further enhances the representation of important feature channels by adaptively adjusting the weight of each channel, enabling effective feature extraction of the amplitude component. This process outputs the feature component The specific process formula is as follows:

[0169] Y′ lum =Conv(Y lum )

[0170] Y" lum =LeakyReLU(Y′ lum )

[0171] ″′ lum =SELayer(Y″) lum )

[0172]

[0173] where Y′ lum , Y″ lum and Y″′ lum is the extracted intermediate feature, It is the characteristic component of the final output.

[0174] Step 2.1.2: Low-light image I low and the enhanced predicted amplitude component Input into the amplitude enhancement module (AmpEnhancer), such as Figure 5As shown in Figure 1, the amplitude enhancement in the Fourier domain is performed, and the predicted amplitude is combined with the low-light image amplitude. This module contains six enhancement blocks, each of which outputs a further enhanced image as the input of the next module. Jump connections are made between the two enhancement modules to combine shallow features with deep features to avoid the loss of detail information. For the low-light image I input to the enhancement module low First, perform Fourier transform to the frequency domain and extract the amplitude component A low and phase component P low :

[0175] A low , P low =FFT(I low )

[0176] Then the two components are respectively subjected to two 1x1 convolutions, LeakyReLU activation function and SELayer for feature extraction to obtain A′ low and P′ low :

[0177] A′ low =Conv(SE(Lea(Conv(A low ))))

[0178] P′ low =Conv(SE(Lea(Conv(P low ))))

[0179] Among them, Conv represents 1x1 convolution, SE represents SELayer, and Lea represents LeakyReLU activation function.

[0180] Step 2.1.3, A′ low Use the sigmoid activation function to limit it to the range of (0,1), and then A′ low and Y′ lum After pixel-by-pixel addition and Sigmoid(A′ low ) pixel by pixel, and then perform residual connection to obtain the enhanced amplitude A that combines the feature extraction amplitude component of the low-light image and the feature extraction amplitude component of the YCbCr space out . Then A out and P′ low After inverse Fourier transform, the output I′ of the enhancement module is obtained low The output is then fed into the next enhancement module to perform the same operation as in step 2.1.3 until all six enhancement modules are completed. This fusion method can effectively restore the image's lack of brightness while also reducing noise interference. The formula for this process is as follows:

[0181]

[0182] I′ low =F -1 (A out , P′ low )

[0183] in, represents pixel-by-pixel addition, Indicates pixel-by-pixel multiplication, F -1 represents the inverse Fourier transform.

[0184] Step 2.2: Input the low-light image into the second branch, the phase enhancement branch that combines Fourier and Transformer, to restore the image detail information and obtain the detail restored image I. out2 .

[0185] The phase information of the Fourier transform contains the structure and details of the image, reflecting the edges and texture structure of the image, and has been proven to be effective in improving image quality. The Transformer architecture, due to its powerful self-attention mechanism, can capture the global information of the image and effectively understand the context and details of the image by modeling long-range dependencies. Therefore, in the second branch, the phase information of the image is introduced into the attention, which can help restore the image's representation and structural information. The specific process is as follows:

[0186] Step 2.2.1, as Figure 4 As shown, the low light image I low (b,c,h,w) is first reshaped to (b,h*w,c) so that the normalization operation can be performed independently on all pixel positions of each channel. The transformed image is then input into the layer normalization module to normalize the channel number dimension. The processed tensor is then reshaped to obtain I ln (b,c,h,w).

[0187] Step 2.2.2, normalize the feature I ln (b,c,h,w) is input into the phase-guided attention module (PGA), as Figure 6 As shown, the feature I after attention weighting is obtained attn (b,c,h,w). By introducing phase information to guide the feature attention mechanism, the network can better focus on important high-frequency information and effectively enhance the perception of local features and key areas such as image details and textures. Specifically, I lnFirst, a convolution operator with a kernel size of 1x1 is used to generate a feature map, expanding the number of channels to three times the original number (b, 3*c, h, w), providing richer expression capabilities for the subsequent attention mechanism. The features are further processed through a 3x3 deep convolution layer. Deep convolution can apply the convolution operation to each channel and capture the relationship between adjacent pixels. The network is then evenly divided into three parts along the dimension of the number of channels, which serve as the query (q(b,c,h,w)), key (k(b,c,h,w)) and value (v(b,c,h,w)) of the attention module.

[0188] kqv=Conv3(Conv1(I in ))

[0189] Among them, Conv1 represents 1x1 convolution and Conv3 represents 3x3 convolution.

[0190] Step 2.2.3: In order to further enhance the structural features of q and k and enhance the phase of the image to obtain the detailed features of the image, the obtained q and k components are respectively input into the phase enhancement module (PB) to extract the detailed contours contained in the phase, as shown in Figure 7 Specifically, in the phase enhancement module, the input q(b,c,h,w) and k(b,c,h,w) are first subjected to a 1x1 convolution to perform preliminary feature adjustment, and then undergo Fourier transform to extract the phase components P of q and k. q and P k :

[0191] P q =FFT(Conv(q))

[0192] P k =FFT(Conv(k))

[0193] In order to further enhance the phase component, the phase component is input into two 1x1 convolutions, LeakyReLU activation function and SELayer layer to obtain the enhanced phase P′ q and P′ k ; Figure 7 The Trans operation of the amplitude branch in the example sets the amplitude component to 1. This operation is to focus only on the increase of the phase branch; then the enhanced phase P′ q and P′ k The amplitude set to 1 is transformed from the frequency domain to the spatial domain through inverse Fourier transform to obtain φ q and φ k To proceed to the next step. The formula of this process is as follows:

[0194] P′ q=Conv(SE(Lea(Conv(P q )))),P′ k =Conv(SE(Lea(Conv(P k ))))

[0195] φ q =iFFT(P′ q ),φ k =iFFT(P′ k )

[0196] Among them, Conv represents 1x1 convolution, SE represents SELayer, Lea represents LeakyReLU activation function, Y q and Y k Represents the generated intermediate variable.

[0197] Step 2.2.4, such as Figure 6 As shown, for the output φ of the phase enhancement module q (b,c,h,w) and φ k (b,c,h,w), and v(b,c,h,w) are reshaped to obtain the shape (b,1,c,h*w). The purpose of this operation is to expand each spatial dimension (h,w) into a long vector, which is convenient for subsequent attention calculation; then L2 normalization is performed to obtain and Make the L2 norm of each of their vectors 1 to ensure that the influence of each position is on the same scale when calculating the attention weight. The formula is as follows:

[0198]

[0199] Among them, PB represents the phase enhancement module and R represents the dimension reshaping process.

[0200] Afterwards and Input to the attention module for context enrichment. The specific process is: first calculate and The similarity between them is calculated by matrix multiplication to obtain an attention weight matrix of shape (b,1,h*w,h*w), which represents the similarity between each pixel. Then the similarity matrix is passed through the softmax function and multiplied by the scaling factor temperature to scale the attention weight. Perform weighted summation and reshape to (b, c, h, w) to obtain the final output I of the phase-guided attention module attn (b,c,h,w). The formula for this process is as follows:

[0201]

[0202] Among them, represents matrix multiplication, Softmax represents the softmax activation function, and temp represents the scaling factor.

[0203] Step 2.2.5: Direct the output of the attention module I attn (b,c,h,w) and I low (b,c,h,w) and I ln (b,c,h,w) are added element by element to obtain the further output Y(b,c,h,w), the formula is as follows:

[0204]

[0205] in, Represents element-wise addition.

[0206] Then, Y(b,c,h,w) is input into the layer normalization module (Layer Normalization) to obtain Y ln , and then Y ln Input into the feedforward neural network FRFN to eliminate redundant features and improve the network's recovery ability to obtain the output Y fn , the formula for this process is as follows:

[0207] Y ln =LN(Y)

[0208] Y fn =FRFN(Y ln )

[0209] Among them, LN represents the layer normalization module and FRFN represents the feedforward neural network.

[0210] Finally, Y, Y ln and Y fn Perform element-by-element addition to obtain the output I of the phase enhancement branch out2 :

[0211]

[0212] Step 2.3: The output result of the first branch amplitude enhancement I out1 and the second branch phase enhancement result I out2 Input into the feature fusion module SKFusion (selective kernel feature fusion) to obtain the final low-light image enhancement result I out , the formula is as follows:

[0213] I out =S K (Iout1 , I out2 )

[0214] The SKFusion module selectively adjusts the size of the receptive field, enabling the model to adaptively capture information at different scales, thereby enhancing the expressive power of feature representation.

[0215] Step 3: Based on the mean square error loss and perceptual loss, a loss function is constructed between the amplitude enhancement branch and the brightness enhancement branch and the normal lighting image. This allows the model to optimize the generated image at multiple levels, ensuring that the final output image is not only close to the normal lighting image at the pixel level, but also has a certain improvement in high-level visual features, texture details, and edge clarity. Specifically, the overall loss function consists of the following five parts:

[0216] L total =L s1 +λ1L vgg +λ2L s2 +λ3L s3 +λ4L lum

[0217] Here, λ1, λ2, λ3, and λ4 represent loss weights, which are used to balance the contributions of different loss terms to training, ensuring that the model can find an appropriate balance between multiple objectives. λ1, λ2, λ3, and λ4 are empirically set to λ1, λ2, λ3, λ4 = [0.3, 0.2, 0.2, 0.1].

[0218] L s1 Indicates the final output I out The L2 loss between the prediction error of each sample is squared to penalize large errors, and then the squared errors of all samples are averaged to ensure that the errors of all samples are comprehensively measured instead of focusing on the errors of individual samples. s1 The formula is as follows:

[0219] L s1 =||I out1 -GT||2

[0220] Among them, GT represents the real normal lighting image.

[0221] L vgg Indicates the final output I out The perceptual loss between L and GT constrains the features extracted from VGG. s2 Represents the output result of the amplitude enhancement branch I out1 The L2 loss between the first branch and the GT enables the network of the first branch to generate images close to natural lighting, improving the amplitude enhancement ability of the branch. The formula is as follows:

[0222] L s2 =||I out1 -GT||2

[0223] L s3 Represents the output result of the phase enhancement branch I out2 The L2 loss between the GT and the second branch is used to optimize the ability of the second branch to enhance the phase so that it can generate an output close to the detail structure of the natural illumination image. The formula is as follows:

[0224] L s3 =||I out2 -GT||2

[0225] L lum Represents the L2 loss between the output of the brightness prediction network (LumiNet) and the Y component of the YCbCr space of the normal light image. This loss can constrain the training of the prediction network to ensure that the generated brightness image is consistent with the Y component of the real image at the pixel level:

[0226] L lum =||LumiNet(Y low )-Y high ||2

[0227] Among them, Y high Represents the Y component of the YCbCr space of the normal light image.

[0228] Step 4: Perform image enhancement on the LOLv2-real benchmark dataset using the optimized low-light image enhancement network model. The LOLv2-real dataset was captured in real-world scenes by adjusting exposure time and ISO values, making it representative of images captured in low-light environments.

[0229] Step 4.1: Convert the low-light image to YCbCr color space to extract the brightness component Y low , and then input it into the amplitude enhancement network branch together with the low-light image to obtain a brightness-enhanced image.

[0230] Step 4.2: Input the low-light image into the phase enhancement network branch to restore detail information and obtain an image with restored texture details.

[0231] Step 4.3: Fuse the features of the images output by the two branches to obtain the final enhanced image and save the image. Figure 8 The figure is a comparison of a low-light image and the generated enhanced image, which reflects the effectiveness of the present invention and can better restore images in low-light environments.

[0232] Through the above-mentioned method, the present invention performs a low-light image enhancement method based on YCbCr space and Fourier frequency domain, and constructs a dual-branch network, an amplitude enhancement network branch and a phase enhancement network branch, so as to simultaneously restore the details and brightness information of the image during the enhancement process, thereby improving the visual quality of the low-light image. In the amplitude enhancement network branch, the brightness component of the YCbCr space is introduced, and the amplitude of the brightness component and the amplitude of the low-light image are enhanced and combined, thereby achieving the effect of improving the brightness of the low-light image. In the phase branch, by enhancing the phase information in the attention module of the Transformer structure, the long-distance dependency between different spatial positions is captured, which helps to restore the edges, textures and other important details of the image. A selective kernel feature fusion module is used after the dual branches to fuse the enhancement results of the two branches to ensure that the detail information and global structure of the image can be fully utilized, further improving the clarity, contrast and detail recovery ability of the image.

Claims

1. A low-light image enhancement method based on YCbCr space and Fourier frequency domain, characterized in that: The following steps are involved: Step 1: Extract the brightness component of the low-light image in the YCbCr space and construct a training set together with the normal-light image and the low-light image; Step 2: Build a low-light image enhancement network model including an amplitude enhancement network branch, a phase enhancement network branch, and a feature fusion module, and input the training set obtained in step 1 into the model for training; Step 3: Based on the mean square error loss and perceptual loss, a loss function is constructed between the amplitude enhancement network branch and the phase enhancement network branch in the model obtained in step 2 and the normal light image. The model parameters are adjusted by minimizing the loss function to obtain an optimized low-light image enhancement network model. Step 4: Input the low-light image to be enhanced into the model obtained in step 3 for image enhancement.

2. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1: The low-light image I in the training set low and paired brightness component Y low The amplitude enhancement network branch is input together to obtain the enhanced image I after brightness improvement. out1 ; Step 2.2: The low-light image I in the training set low The input is fed into two serial phase enhancement network branches. The output of the first branch is fed into the input of the second branch, and the enhanced image I with restored detail structure information is obtained. out2 ; Step 2.3: Image I obtained in step 2.1 out1 and image I obtained in step 2.2 out2 Input feature fusion module, adaptively fuse the different features of the two branches, and obtain the enhanced image I out .

3. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 2, characterized in that: The step 2.1 specifically includes the following steps: Step 2.1.1: Convert the brightness component Y of the low-light image low Input the brightness prediction network LumiNet to predict the Y component of the normal lighting image in the YCbCr space and obtain the predicted component Y′: Y′=LumiNet(Y low ) The predicted component Y′ is then converted to the Fourier domain through Fourier transform, and the amplitude component Y in the Fourier domain is extracted. lum , the Fourier transform formula is: Among them, X(u,v) is the complex component, H and W are the height and width of the input image, h and w are the coordinates in the spatial domain, u and v are the coordinates in Fourier space, j is the imaginary unit, and the inverse Fourier transform process is represented by F -1 ; Each complex component X(u,v) is represented by the amplitude component A(X(u,v)) and the phase component P(X(u,v)) as follows: Where R(X) and I(X) represent the real and imaginary parts of X(u,v), respectively; For the amplitude component Y lum After two 1x1 convolutions, LeakyReLU activation function and SELayer feature extraction, the feature components are obtained The specific process formula is as follows: AND' lum =Conv(Y lum ) AND" lum =LeakyReLU(Y′ lum ) AND"' lum =SELayer(Y″ lum ) Among them, Y′ lum , Y″ lum and Y″′ lum is the extracted intermediate feature, is the characteristic component of the final output; Step 2.1.2: Low-light image I low and characteristic components The input is sent to the amplitude enhancement module AmpEnhancer for amplitude enhancement in the Fourier domain. The amplitude enhancement module AmpEnhancer contains six enhancement blocks. Each enhancement block outputs a further enhanced image as the input of the next enhancement block. Jump connections are made between the two enhancement blocks. For the low-light image I input to the enhancement block low First, perform Fourier transform to the frequency domain and extract the amplitude component A low and phase component P low , and then component A low and P low Feature extraction is performed through two 1x1 convolutions, LeakyReLU activation function and SELayer to obtain A′ low and P′ low ; Step 2.1.3: Component A′ low Use the sigmoid activation function to limit A′ to the range of (0,1). low and Y′ lum After pixel-by-pixel addition and Sigmoid(A′ low ) pixel by pixel, and then perform residual connection to obtain the enhanced amplitude A that combines the feature extraction amplitude component of the low-light image and the feature extraction amplitude component of the YCbCr space out , then A out and P′ low After inverse Fourier transform, the output I′ of the enhancement block is obtained low : I′ low =F -1 (A out ,P′ low ) in, represents pixel-by-pixel addition, Indicates pixel-by-pixel multiplication, F -1 represents the inverse Fourier transform; Finally, the output I′ of the current enhancement block is low Input to the next enhancement block and perform the same operation until the six enhancement blocks are completed, and the enhanced image I with improved brightness is obtained. out1 .

4. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 2, characterized in that: The step 2.2 specifically includes the following steps: Step 2.2.1: First, the low-light image I low (b,c,h,w) is reshaped, and the image dimension is transformed into (b,h*w,c). Then the transformed image is input into the layer normalization module to normalize the channel number dimension. The processed tensor is then restored to its shape to obtain I ln (b,c,h,w); Step 2.2.2, normalize the feature I ln (b,c,h,w) is input into the phase-guided attention module PGA to obtain the attention-weighted feature I attn (b,c,h,w); specifically: I ln (b,c,h,w) first passes through a convolution operator with a kernel size of 1x1 to generate a feature map, and the number of channels is expanded by 3 times to obtain (b,3*c,h,w). Then, it passes through a 3x3 deep convolution layer to further process the features. Then, it is evenly split into three parts along the dimension of the number of channels, which are used as the query (q(b,c,h,w)), key (k(b,c,h,w)) and value (v(b,c,h,w)) of the attention module respectively: k,q,v=Conv3(Conv1(I in )) Among them, Conv1 represents 1x1 convolution, and Conv3 represents 3x3 convolution; Step 2.2.3: Input the obtained q and k components into the phase enhancement module PB to extract the detailed outline contained in the phase; specifically: in the phase enhancement module, the input q(b,c,h,w) and k(b,c,h,w) are first subjected to a 1x1 convolution to perform preliminary feature adjustment, and then the phase components P of q and k are extracted through Fourier transform. q and P k , and then the phase component P q and P k Input into two 1x1 convolutions, LeakyReLU activation function and SELayer layer to get the enhanced phase P′ q and P′ k , and then the enhanced phase and the amplitude set to 1 are transformed from the frequency domain to the spatial domain through inverse Fourier transform to obtain φ q and φ k : P q =FFT(Conv(q)),P k =FFT(Conv(k)) P′ q =Conv(E(Lea(Conv(P q ))),P′ k =Conv(SE(Lea(Conv(P k )))) f q =iFFT(P′ q ),φ k =iFFT(P′ k ) Among them, Conv represents 1x1 convolution, SE represents SELayer, Lea represents LeakyReLU activation function, Y q and Y k Represents the generated intermediate variables; Step 2.2.4: Output φ of the phase enhancement module q (b,c,h,w) and φ k (b,c,h,w), and v(b,c,h,w) are reshaped and L2 normalized to become and Make the L2 norm of each vector 1, and then and Input to the attention module for context enrichment, specifically: first calculate and The similarity between them is measured by matrix multiplication to obtain an attention weight matrix of shape (b, 1, h*w, h*w), which represents the similarity between each pixel. Then the similarity matrix is passed through the softmax function and multiplied by the scaling factor temperature to scale the attention weight. Perform weighted summation and reshape to (b, c, h, w) to obtain the final output I of the phase-guided attention module attn (b,c,h,w): Among them, PB represents the phase enhancement module, R represents the dimension reshaping process, represents matrix multiplication, and temp represents the scaling factor; Step 2.2.5: Direct the output of the attention module I attn (b,c,h,w),I low (b,c,h,w) and I ln (b,c,h,w) are added to get the output Y(b,c,h,w): Then input Y(b,c,h,w) into the layer normalization module to obtain Y ln , and then Y ln Input into the feedforward neural network FRFN to get the output Y fn : Y ln =LN(Y) Y fn =FRFN(Y ln ) Finally, Y, Y ln and Y fn Perform element-by-element addition to obtain the output I of the second phase enhancement network branch out2 :

5. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 2, characterized in that: The step 2.3 is specifically as follows: the image I obtained in step 2.1 is out1 and image I obtained in step 2.2 out2 Input the feature fusion module SKFusion to obtain the enhanced image I out : I out =SK(I out1 ,I out2 )。 6. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 5, characterized in that: The loss function constructed in step 3 is expressed as: L total =L s1 +λ1L vgg +λ2L s2 +λ3L s3 +λ4L lum Among them, λ1, λ2, λ3 and λ4 represent loss weights; L s1 Represents the enhanced image I output by the low-light image enhancement network model out And the L2 loss between the real normal lighting image GT: L s1 =||I out -GT||2 L vgg Represents the enhanced image I output by the low-light image enhancement network model out The perceptual loss between the real normal lighting image GT constrains the features extracted from VGG; L s2 represents the output of the amplitude enhancement network branch I out1 And the L2 loss between the real normal lighting image GT: L s2 =||I out1 -GT||2 L s3 I represents the output of the phase enhancement network branch out2 And the L2 loss between the real normal lighting image GT: L s3 =||I out2 -GT||2 L lum Represents the output of the brightness prediction network LumiNet and the Y component Y of the normal lighting image YCbCr space high The L2 loss between: L lum =||LumiNet(Y low )-AND high ||2。 7. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 6, characterized in that: The loss weights of the loss function are set to λ1, λ2, λ3, λ4 = [0.3, 0.2, 0.2, 0.1].

8. The low-light image enhancement method based on YCbCr space and Fourier frequency domain according to claim 1, characterized in that: The step 4 specifically includes the following steps: Step 4.1: First, convert the low-light image to be enhanced into YCbCr space to extract the brightness component Y low , and then input it into the amplitude enhancement network branch together with the low-light image to obtain a brightness-enhanced image; Step 4.2: Input the low-light image to be enhanced into the phase enhancement network branch to restore the detail structure information and obtain an image with restored texture details; Step 4.3: Feature fusion is performed on the images output by the amplitude enhancement network branch and the phase enhancement network branch to obtain an enhanced image.

Images