Weak light image enhancement method based on YCbCr brightness guidance

Abstract

The invention discloses a weak light image enhancement method based on YCbCr brightness guidance, and the method specifically comprises the steps: converting a normal light image and a weak light image from an RGB color space to a YCbCr color space, obtaining a brightness component, a blue component and a red component, inputting all the brightness components into a brightness component diffusion enhancement branch, obtaining an enhanced component, enabling the enhanced brightness component to be subjected to the YCbCr brightness guidance, and enabling the enhanced brightness component to be subjected to the YCbCr brightness guidance to be subjected to the YCbCr brightness guidance. The blue component, the red component and the enhanced component of the weak light image are spliced and then input into a chrominance component recovery branch to obtain an enhanced blue component and an enhanced red component, the enhanced brightness component, the enhanced blue component and the enhanced red component are converted into an RGB color space to obtain an enhanced image, and loss constraint is carried out to obtain an enhanced image; and a final enhanced image is obtained. According to the method, the problems of poor color and brightness recovery effect and unclear enhancement result texture of the existing method are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image enhancement methods, and particularly relates to a weak light image enhancement method based on YCbCr brightness guidance. BACKGROUND

[0002] In actual application, in scenes such as night photography, security monitoring, vehicle night vision, etc., images often present problems such as low brightness, detail loss, color distortion, etc. due to insufficient illumination, which seriously affects subsequent visual understanding and perception tasks. Traditional image enhancement methods mainly include histogram equalization, Retinex model, Gamma correction, etc. These methods can improve brightness to a certain extent, but when dealing with extremely low light environments, they often cannot effectively restore details, and are prone to cause image overexposure or noise amplification.

[0003] In recent years, deep learning technology has been widely applied to low-illumination image enhancement tasks. Convolutional neural networks (CNN) and deep models based on Retinex decomposition have gradually become mainstream, but there are the following shortcomings: (1) brightness and color restoration are not modeled separately, resulting in poor color and brightness restoration; (2) lack of modeling of image structure and frequency components, resulting in unclear texture in the enhancement result.

[0004] Therefore, there is an urgent need for a weak light image enhancement method with strong restoration capability and clear texture. SUMMARY

[0005] The purpose of the present application is to provide a weak light image enhancement method based on YCbCr brightness guidance, which solves the problems of poor color and brightness restoration and unclear texture in the enhancement result of existing methods.

[0006] The technical solution adopted by the present application is a weak light image enhancement method based on YCbCr brightness guidance, and the specific process is as follows: normal light images and weak light images are converted from RGB color space to YCbCr color space to obtain brightness components, blue components and red components; all brightness components are input into a brightness component diffusion enhancement branch to obtain enhanced components and enhanced brightness components; the blue components and red components of the weak light image and the enhanced components are spliced and input into a chroma component restoration branch to obtain enhanced blue components and enhanced red components; the enhanced brightness components, enhanced blue components and enhanced red components are converted to RGB color space to obtain enhanced images, and loss constraints are performed to obtain the final enhanced images. The present application has the following characteristics: The specific implementation steps are as follows: Step 1, convert the normal light image from RGB color space to YCbCr color space to obtain brightness components , blue components , red components The weak light image is converted from the RGB color space to the YCbCr color space to obtain a luminance component , a blue component , and a red component ; Step 2, the luminance component , the luminance component is input into a luminance component diffusion enhancement branch to obtain an enhanced component , an enhanced luminance component ; Step 3, the blue component , the red component , and the enhanced component are spliced in the channel dimension and then input into a chrominance component recovery branch to obtain an enhanced blue component , and an enhanced red component ; Step 4, the enhanced luminance component , the enhanced blue component , and the enhanced red component are converted into the RGB color space to obtain an enhanced image , and are constrained by a frequency perception loss to obtain a final enhanced image.

[0007] In step 1, the expressions for converting the normal light image and the weak light image from the RGB color space to the YCbCr color space are the same, and the following expression is used uniformly: (1) In formula (1), Y is the luminance component, Cb is the blue component, Cr is the red component, R is the red channel, G is the green channel, B is the blue channel.

[0008] The specific process of step 2 is as follows: Step 2.1, the luminance component , the luminance component is input into a luminance representation encoder, respectively, to output latent feature representations , and latent feature representations ; Step 2.2, the luminance structural noise is sampled from a luminance structure noise table , and the luminance structural noise is used to add noise to the latent feature representations to obtain noisy latent feature representations the noisy latent feature representation with the latent feature representation performing a concatnation operation, and then inputting into a U-Net network to perform prediction to obtain a predicted noise , and obtaining an enhanced component after several update iterations ; Step 2.3, inputting the obtained enhanced component into a decoder to obtain an enhanced luminance component .

[0009] In step 2.1, the luminance representation encoder is sequentially composed of a two-dimensional convolution layer, a first ReLU activation function, five first residual blocks, a first convolution fusion layer, a global average pooling, a first linear layer, a second ReLU activation function, a second linear layer, and a third ReLU activation function.

[0010] Each first residual block is sequentially composed of a first convolution layer, a fourth ReLU activation function, and a second convolution layer, and the input of the first convolution layer and the output of the second convolution layer are element-wise added to construct a residual connection.

[0011] In step 2.2, the expression of the noisy representation is: (2) In formula (2), is the luminance structural noise at the t-th step, which is obtained by sampling from a luminance structural noise table ; is the noisy latent feature representation at the t-th step; wherein, (3) (4) In formulas (3)-(4), is the cumulative residual factor from the first step to the t-th step; is the noise amplitude at the t-th step, which is located between 0 and 1; The expression of the luminance structural noise table is: (5) In formula (5), (x, y) is the pixel coordinate of the image; is the reserved proportion; T is the total number of forward noise addition; is the region of "insufficient brightness"; , , is the weighting coefficient between the three terms; is the local gradient; wherein, (6) In formula (6), s is an offset, which prevents the occurrence of division by zero operation.

[0012] In step 2.3, the decoder is sequentially composed of a third linear layer, a fifth ReLU activation function, a fourth linear layer, a sixth activation function, an up-sampling operation, a second convolution fusion layer, five second residual blocks, a seventh activation function, and an output convolution layer; Each second residual block is sequentially composed of a third convolution layer, an eighth ReLU activation function, and a fourth convolution layer, and the input of the third convolution layer and the output of the fourth convolution layer are element-wise added to construct a residual connection.

[0013] In step 3, the chroma component recovery branch is composed of a first 2D convolution layer, a first KAN-Block, a first DownSample module, a second KAN-Block, a second DownSample module, a third KAN-Block, a third DownSample module, a fourth KAN-Block, a fifth KAN-Block, a first UpSample module, a sixth KAN-Block, a second UpSample module, a seventh KAN-Block, a third UpSample module, an eighth KAN-Block, and a second 2D convolution layer; the blue component , the red component , and the enhanced component After being spliced in the channel dimension, the first 2D convolution layer is taken as the input, the output of the first KAN-Block and the output channel of the third UpSample module are spliced to be taken as the input of the eighth KAN-Block, the output of the second KAN-Block and the output channel of the second UpSample module are spliced to be taken as the input of the seventh KAN-Block, and the output of the third KAN-Block and the output channel of the first UpSample module are spliced to be taken as the input of the sixth KAN-Block; Each KAN-Block has the same structure and is sequentially composed of a first KANLayer, a first DwConv, a second KANLayer, a second DwConv, a third KANLayer, and a third DwConv; The input of each KAN-Block is the output feature of the previous stage and the enhanced component after fusion , and the expression of the fusion is:

[0014] wherein, denotes a Sigmoid activation function.

[0015] In step 2, the loss function of the component diffusion enhancement branch is: (7) In formula (7), is a noise prediction loss; is a luminance reconstruction loss; is a predicted noise; is a luminance structure noise at the t-th step; is a luminance image output by the decoder; In step 3, the loss function of the chrominance component recovery branch is: (8).

[0016] In step 4, the expression of the frequency-aware loss is: (11) In formula (11), is a normal light image; FFT is a fast Fourier transform, is a phase and amplitude of the enhanced image, is a phase and amplitude of the normal light image, , are hyperparameters, respectively.

[0017] The beneficial effects of the present application are: (1) The low-light image enhancement method based on YCbCr luminance guidance of the present application converts the image from the RGB space to the YCbCr space, decouples the luminance and chrominance, constructs a luminance component diffusion enhancement branch and a chrominance component recovery branch in the luminance enhancement and color recovery processes, respectively, and improves the image detail and color recovery fidelity of the enhancement result; (2) The low-light image enhancement method based on YCbCr luminance guidance of the present application introduces a diffusion modeling mechanism performed in a feature latent space in the constructed luminance component diffusion enhancement branch, and combines a structure-aware noise scheduling strategy, thereby enhancing the noise resistance of image detail and texture recovery; (3) The low-light image enhancement method based on YCbCr luminance guidance of the present application uses a learnable activation function and a spline connection weight to improve the modeling capability of nonlinear color relationships in the constructed chrominance component recovery branch, solves the problems of low saturation and color distortion caused by the weak light enhancement recovery process of existing methods, and introduces luminance structure features as a regulatory factor at multiple levels to improve the consistency and structure-aware capability of color recovery; (4) The weak light image enhancement method based on YCbCr brightness guidance of the application, through the loss function constructed, realizes multi-dimensional collaborative enhancement of image brightness, color and perceptual quality. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flow chart of the weak light image enhancement method based on YCbCr brightness guidance of the application; Figure 2 The flow chart of the brightness component diffusion enhancement branch in the weak light image enhancement method based on YCbCr brightness guidance of the application; Figure 3 The structural schematic diagram of the brightness representation encoder in the weak light image enhancement method based on YCbCr brightness guidance of the application; Figure 4 The structural schematic diagram of the chroma component recovery branch in the weak light image enhancement method based on YCbCr brightness guidance of the application. DETAILED DESCRIPTION

[0019] The application will be described in detail below in combination with the drawings and specific embodiments.

[0020] Embodiment 1 The weak light image enhancement method based on YCbCr brightness guidance of the application, the specific process is as follows: the normal light image and the weak light image are both converted from the RGB color space to the YCbCr color space, the brightness component, the blue component and the red component are obtained, all the brightness components are input into the brightness component diffusion enhancement branch, the enhanced component and the enhanced brightness component are obtained, the blue component and the red component of the weak light image and the enhanced component are spliced and input into the chroma component recovery branch, the enhanced blue component and the enhanced red component are obtained, the enhanced brightness component, the enhanced blue component and the enhanced red component are converted into the RGB color space, the enhanced image is obtained, and the loss constraint is performed to obtain the final enhanced image.

[0021] Embodiment 2 The weak light image enhancement method based on YCbCr brightness guidance of the application, as shown in the figure, the specific steps are as follows: Figure 1 Step 1, the normal light image is converted from the RGB color space to the YCbCr color space, the brightness component , the blue component , and the red component are obtained. The weak light image is converted from the RGB color space to the YCbCr color space, the brightness component , the blue component , and the red component are obtained. The conversion formula of the normal light image and the weak light image is the same, and is expressed as follows: (1) In formula (1), Y is a luminance component, Cb is a blue component, Cr is a red component, R is a red channel, G is a green channel, B is a blue channel; Step 2, the luminance component , the luminance component are input into a luminance component diffusion enhancement branch to obtain an enhanced component , an enhanced luminance component ; As shown in Figure 2 , the processing procedure of the component diffusion enhancement branch is as follows: Step 2.1, the luminance component , the luminance component are input into a luminance representation encoder to output latent feature representations , latent feature representation ; The sizes of the luminance component , the luminance component are both 600x400x1; Step 2.2, luminance structure noise is sampled from a luminance structure noise table , and the latent feature representation is added with the luminance structure noise to obtain an added latent feature representation , and the added latent feature representation is concatenated with the latent feature representation , and then input into a U-Net network for prediction to obtain predicted noise , which gradually approaches the added real noise , and after several update iterations, the enhanced component is obtained; The added expression is as follows: (2) In formula (2), is the latent feature representation ; is the luminance structure noise at the tth step, which is sampled from a luminance structure noise table ; is the added latent feature representation at the tth step; wherein, (3) (4) (5) In formula (5), (x, y) is the pixel coordinate of the image; x, y is the noise amplitude of the t-th step, which is located between [0, 1]; is the reserved proportion; is the cumulative residual factor from the first step to the t-th step; T is the total number of forward noise addition; is the area of “insufficient brightness”, the greater the value, the higher the noise intensity; , , , are the weighting coefficients among the three terms, all of which are defaulted to 1; is the local gradient, when the local brightness structure of the image is obvious (such as the edge or texture brightness structure is obvious), is larger, the disturbance is stronger, and the modeling ability of the key structure area is improved, in the flat area, is close to , the low disturbance is maintained, and the original information is avoided from being excessively damaged; (6) In formula (6), s is the offset, which prevents the occurrence of division by 0 operation; Step 2.3, input the obtained enhancement component Y' into the decoder to obtain the enhanced luminance component Y; ; Step 3, after splicing the blue component B, the red component R and the enhancement component Y in the channel dimension, input them into the chroma component recovery branch to obtain the enhanced blue component B' and the enhanced red component R'; ; The loss function of the chroma component recovery branch is: (8) Step 4, convert the enhanced luminance component Y', the enhanced blue component B' and the enhanced red component R' into the RGB color space to obtain the enhanced image I', and constrain it by the frequency perception loss to obtain the final enhanced image I; The expression of the enhanced image I' is: ​​​​​​​​​​​ (9) In equation (7), This is the converted red channel; This is the converted green channel; This is the converted blue channel.

[0022] Example 3 Based on Example 2, in step 2.1, as follows: Figure 3 As shown, the brightness encoder consists of a two-dimensional convolutional layer, a first ReLU activation function, five first residual blocks, a first convolutional fusion layer, global average pooling, a first linear layer, a second ReLU activation function, a second linear layer, and a third ReLU activation function. The two-dimensional convolutional layer has a 3×3 kernel, 64 channels, a stride of 1, and padding of 1. It is used to extract low-level local brightness features, and the output feature size is 600×400×64. The first ReLU activation function introduces non-linear transformation capabilities, enhancing the modeling effect on complex structures such as image edges and textures; Each first residual block consists of a first convolutional layer, a fourth ReLU activation function, and a second convolutional layer in sequence. The input of the first convolutional layer and the output of the second convolutional layer are added element-wise to construct residual connections. The first residual block is used to extract mid-to-high-level structural information of the image layer by layer, maintain the feature size of 600×400×64, and effectively enhance gradient fluidity and suppress feature degradation. The first convolutional fusion layer takes a feature map of size 600×400×64 as input, uses a 3×3 kernel, has 64 input channels and 64 output channels, a stride of 1, and padding of 1. This layer is used to further integrate local contextual information from the mid-layer brightness features, while maintaining the output size of 600×400×64.

[0023] Global average pooling compresses the spatial dimension to 1×1, resulting in a channel feature vector of size 1×1×64, which represents the global structure representation of the entire image. The first linear layer expands the channel dimension from 64 to 128; The second linear layer maps the dimension to the final set latent dimension of 256, and outputs a latent feature representation of size 1×1×256.

[0024] Example 4 Based on Example 3, in step 2.3, the decoder consists of a third linear layer, a fifth ReLU activation function, a fourth linear layer, a sixth activation function, an upsampling operation, a second convolutional fusion layer, five second residual blocks, a seventh activation function, and an output convolutional layer. The third linear layer reduces the enhanced component The fourth linear layer reduces the 128 dimensions to 64 dimensions, the fifth ReLU activation function and the sixth activation function introduce the nonlinear transformation ability, and the output feature size of the sixth activation function is still 1x1x64; The up-sampling operation up-samples the 1x1x64 feature tensor to 600x400x64, and restores it to the image space dimension; The convolution kernel of the second convolution fusion layer is 3x3, the channel number is 64, the step is 1, and the padding is 1, which is used to preliminarily integrate the up-sampling feature, and the output size is kept as 600x400x64; Each second residual block is composed of a third convolution layer, an eighth ReLU activation function, and a fourth convolution layer in turn, and the input and output of each second residual block are added element by element to construct a residual connection; the second residual block is used to enhance the middle layer structure recovery ability of the image, and the feature size does not change; The convolution kernel of the output convolution layer is 3x3, the input channel is 64, and the output channel is 1, and the size of the output enhanced luminance component is 600x400x1.

[0025] Example 5 On the basis of example 4, in step 3, as Figure 4 ​As shown, the chroma component recovery branch is composed of a first 2D convolution layer, a first KAN-Block, a first DownSample module, a second KAN-Block, a second DownSample module, a third KAN-Block, a third DownSample module, a fourth KAN-Block, a fifth KAN-Block, a first UpSample module, a sixth KAN-Block, a second UpSample module, a seventh KAN-Block, a third UpSample module, an eighth KAN-Block, and a second 2D convolution layer; the output of the first KAN-Block and the output channel of the third UpSample module are spliced to serve as the input of the eighth KAN-Block, the output of the second KAN-Block and the output channel of the second UpSample module are spliced to serve as the input of the seventh KAN-Block, and the output of the third KAN-Block and the output channel of the first UpSample module are spliced to serve as the input of the sixth KAN-Block; wherein the first 2D convolution layer adopts a 3x3 convolution kernel, the number of channels is set to 64, the step is 1, and the padding is 1; the three DownSample modules are all realized by a standard convolution layer with a 3x3 convolution kernel, a step of 2, a padding of 1, and a number of channels of 64 to perform down-sampling operation; the three UpSample modules are all realized by a bilinear interpolation method to restore the feature map to the spatial resolution of the previous level; the second 2D convolution layer adopts a 3x3 convolution kernel, the input channel number is 64, the output channel number is 2, the step is 1, and the padding is 1, and is used for outputting the enhanced blue component , the enhanced red component .

[0026] The structure of each KAN-Block is the same, and is sequentially composed of a first KANLayer, a first DwConv, a second KANLayer, a second DwConv, a third KANLayer, and a third DwConv; has a learnable activation function and a spline weight representation capability, and effectively adapts to the highly nonlinear and locally discontinuous distribution in color mapping; wherein the three DwConvs are each composed of a 3x3 channel-by-channel convolution and a 1x1 point-by-point convolution, the number of channels is 64, the step is 1, and the padding is 1, and are used for enhancing the feature expression capability and the network nonlinear modeling capability; The first KAN-Block, the second KAN-Block and the third KAN-Block are used to extract and enhance the nonlinear relationship features between chroma and structure; the first DownSample module, the second DownSample module and the third DownSample module realize spatial compression and receptive field expansion; the fourth KAN-Block and the fifth KAN-Block further extract deep semantic features and enhance the expression ability for complex color drift and saturation compression problems; The processing process of the chroma component recovery branch is as follows: The blue component , the red component , the enhanced component After being spliced in the channel dimension, the first 2D convolution layer (input size HxWxC, C=3) is input, and the output (size HxWx64) of the first 2D convolution layer and the enhanced component The fused features are taken as the input of the first KAN-Block, and the output of the first KAN-Block is taken as the input of the first DownSample module; the output (size H / 2xW / 2) of the first DownSample module and the enhanced component The fused features are taken as the input of the second KAN-Block, and the output of the second KAN-Block is taken as the input of the second DownSample module; the output (size H / 4xW / 4) of the second DownSample module and the enhanced component The fused features are taken as the input of the third KAN-Block, and the output of the third KAN-Block is taken as the input of the third DownSample module; the output (size H / 8xW / 8) of the third DownSample module is taken as the input of the fourth KAN-Block, the output of the fourth KAN-Block is taken as the input of the fifth KAN-Block, the output of the fifth KAN-Block is taken as the input of the first UpSample module, the output of the third KAN-Block and the output of the first UpSample module are spliced in the channel dimension and taken as the input of the sixth KAN-Block, the output of the sixth KAN-Block is taken as the input of the second UpSample module, the output of the second KAN-Block and the output of the second UpSample module are spliced in the channel dimension and taken as the input of the seventh KAN-Block, the output of the seventh KAN-Block is taken as the input of the third UpSample module, the output of the first KAN-Block and the output of the third UpSample module are spliced in the channel dimension and taken as the input of the eighth KAN-Block, and the output of the eighth KAN-Block is taken as the input of the second 2D convolution layer; the output of the second 2D convolution layer is the enhanced blue component , the enhanced red component (size H x W x 2); The input of each KAN-Block is the output feature of the previous stage and the enhanced component The fused feature The fusion mode is a structural attention regulation mechanism, and its expression is:

[0027] wherein, represents a Sigmoid activation function, and the attention map is used for dynamically adjusting the enhancement intensity of each spatial position in the chroma feature; the structural attention regulation mechanism effectively improves the spatial consistency and structural perception ability in the chroma restoration process by introducing the brightness structural information, and suppresses the color distortion and other problems in the low-light image.

[0028] Embodiment 6 On the basis of embodiment 5, in step 2, the loss function of the component diffusion enhanced branch is: (7) In formula (7), is a noise prediction loss; is a brightness reconstruction loss; is a predicted noise; is the brightness structural noise at the t-th step; is the brightness image output by the decoder; In step 3, the loss function of the chroma component restoration branch is: (8); In step 4, the expression of the frequency perception loss is: (11) In formula (11), is a normal light image; FFT is a fast Fourier transform, is the phase and amplitude of the enhanced image, is the phase and amplitude of the normal light image, , are hyperparameters, and both are set to 0.5 by default.

[0029] The YCbCr brightness guided low-light image enhancement method has good application prospect and practical value, and can realize high-quality enhancement of images in low-light scenes such as night imaging, security, and automatic driving.

Claims

1. A low-light image enhancement method based on YCbCr luminance guide, characterized in that, The specific process is that the normal light image and the weak light image are converted from the RGB color space to the YCbCr color space, the luminance component, the blue component and the red component are obtained, all the luminance components are input into the luminance component diffusion enhancement branch, the enhanced component and the enhanced luminance component are obtained, the blue component and the red component of the weak light image and the enhanced component are spliced and input into the chroma component recovery branch, the enhanced blue component and the enhanced red component are obtained, the enhanced luminance component, the enhanced blue component and the enhanced red component are converted into the RGB color space, the enhanced image is obtained, and loss constraint is carried out, and the final enhanced image is obtained.

2. The YCbCr luma-guided low-light image enhancement method of claim 1, wherein, The specific implementation is as follows: Step 1, converting normal light image from RGB color space to YCbCr color space to get luminance component , blue component , red component , converting weak light image from RGB color space to YCbCr color space to get luminance component , blue component , red component ; Step 2, the luminance component , the luminance component is input into a luminance component diffusion enhancement branch to obtain an enhanced component , the enhanced luminance component ; Step 3, blue component , red component , enhanced component After splicing in the channel dimension, input into the chroma component recovery branch to obtain the enhanced blue component , enhanced red component ; Step 4, the enhanced luminance component , enhanced blue component , enhanced red component is converted to the RGB color space to obtain an enhanced image , and is constrained by a frequency-aware loss to obtain a final enhanced image.

3. The YCbCr luma-guided low-light image enhancement method of claim 2, wherein, In step 1, the expressions of the normal light image and the weak light image converted from the RGB color space to the YCbCr color space are the same, and the following expression is used: （1） In formula (1), Y is a luminance component, Cb is a blue component, Cr is a red component, R is a red channel, G is a green channel, B is a blue channel.

4. The YCbCr luma-guided low-light image enhancement method of claim 2, wherein, The specific process of step 2 is as follows: Step 2.1, the luminance component , the luminance component is input into a luminance representation encoder, outputting a latent feature representation , the latent feature representation ; Step 2.2, from the luminance structural noise table Sampling the luminance structural noise And using the luminance structural noise To the latent feature representation Add noise to obtain the noisy latent feature representation , the noisy latent feature representation And the latent feature representation Concatenate operation, and then input into the U-Net network for prediction to obtain the predicted noise , after several update iterations to obtain the enhancement component ; Step 2.

3. The resulting enhancement component is input into the decoder to obtain the enhanced luminance component .

5. The YCbCr luma-guided low-light image enhancement method according to claim 4, characterized in that, In step 2.1, the luminance representation encoder is composed of a two-dimensional convolution layer, a first ReLU activation function, five first residual blocks, a first convolution fusion layer, global average pooling, a first linear layer, a second ReLU activation function, a second linear layer, a third ReLU activation function in sequence. Each first residual block is composed of a first convolution layer, a fourth ReLU activation function and a second convolution layer in sequence, and the input of the first convolution layer and the output of the second convolution layer are added element by element to construct a residual connection.

6. The YCbCr luma-guided low-light image enhancement method of claim 4, wherein, In step 2.2, the expression of the added noise is as follows: （2） In formula (2), is the luminance structural noise at the t-th step, from the luminance structural noise table is obtained by sampling; is the noisy latent feature representation at the t-th step. Wherein, （3） （4） In formulas (3) to (4), is the residual factor accumulated from step 1 to step t; is the noise amplitude of step t, located between [0, 1]; Luminance structure noise table The expression is: （5） In formula (5), (x, y) is a pixel coordinate of the image; is the reserved ratio; T is the total number of forward noise addition; is a region of "insufficient brightness"; 、 、 is a weighting coefficient among the three terms; is a local gradient; Wherein, （6） In formula (6), s is an offset to prevent division by zero.

7. The YCbCr luma-guided low-light image enhancement method of claim 4, wherein, In step 2.3, the decoder is composed of a third linear layer, a fifth ReLU activation function, a fourth linear layer, a sixth activation function, an up-sampling operation, a second convolution fusion layer, five second residual blocks, a seventh activation function and an output convolution layer in sequence. Each second residual block is composed of a third convolution layer, an eighth ReLU activation function and a fourth convolution layer in sequence, and the input of the third convolution layer and the output of the fourth convolution layer are added element by element to construct a residual connection.

8. The YCbCr luma-guided low-light image enhancement method of claim 2, wherein, In step 3, the chroma component recovery branch is composed of a first 2D convolutional layer, a first KAN-Block, a first DownSample module, a second KAN-Block, a second DownSample module, a third KAN-Block, a third DownSample module, a fourth KAN-Block, a fifth KAN-Block, a first UpSample module, a sixth KAN-Block, a second UpSample module, a seventh KAN-Block, a third UpSample module, an eighth KAN-Block, and a second 2D convolutional layer; the blue component , the red component , and the enhancement component After being spliced in the channel dimension, the output of the first 2D convolutional layer is taken as the input of the first KAN-Block, the output of the third UpSample module is taken as the input of the eighth KAN-Block after being spliced with the output channel of the first KAN-Block, the output of the second UpSample module is taken as the input of the seventh KAN-Block after being spliced with the output channel of the second KAN-Block, and the output of the first UpSample module is taken as the input of the sixth KAN-Block after being spliced with the output channel of the third KAN-Block. Wherein, the structure of each KAN-Block is the same, and is composed of a first KANLayer, a first DwConv, a second KANLayer, a second DwConv, a third KANLayer and a third DwConv in sequence. The input of each KAN-Block is the output feature of the previous stage with the enhanced component The fused feature , whose expression is wherein, denotes a Sigmoid activation function.

9. The YCbCr luma-guided low-light image enhancement method of claim 2, wherein, In step 2, the loss function of the component diffusion augmented branch is: （7） In formula (7), is a noise prediction loss; is a luminance reconstruction loss; is a predicted noise; is a luminance structural noise at the t-th step; is a decoded luminance image; In step 3, the loss function of the chroma component recovery branch is: （8）。 10. The YCbCr luma-guided low-light image enhancement method of claim 2, wherein, In step 4, the expression of the frequency perception loss is as follows: （11） In formula (11), is a normal light image; FFT stands for Fast Fourier Transform. To enhance the phase and amplitude of the image, The phase and amplitude of a normal light image. , These are hyperparameters.

Images