Weak light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception

Through the soft-gated fusion mechanism and adaptive frequency domain perception technology, combined with the frequency domain perception and detail retention module, the problems of texture details loss and noise amplification in image enhancement are solved, which significantly improves the visual quality of low-light images and realizes the image Natural texture retention and visual performance improvement in low-illumination environments.

CN118982472AActive Publication Date: 2024-11-19XIAN UNIV OF TECH
View PDF 0 Cites 14 Cited by

Patent Information

Application Number
CN202411050045.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-19
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

Existing image enhancement technologies are often accompanied by problems such as loss of texture details, noise amplification and color distortion in improving image quality.

Method used

Using a method based on soft-gated fusion mechanism and adaptive frequency domain perception, the frequency domain perception and detail retention module of the diffusion model is analyzed in combination with wavelet transform, the frequency domain information is adaptively processed, and the soft gate is used in the fusion stage. The control mechanism optimizes the interaction between appearance and detail features, ensuring the retention of image natural texture and the improvement of visual quality.

Benefits of technology

It significantly improves the visual quality of low-light images, avoids texture details loss and noise amplification, ensures the natural performance of the image in low-illumination environments, and solves common problems in image enhancement.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982472A_ABST
    Figure CN118982472A_ABST
Patent Text Reader

Abstract

The invention discloses a weak light image enhancement method based on a soft gating fusion mechanism and adaptive frequency domain perception. Designing a frequency domain sensing module based on a diffusion model to carry out frequency domain analysis on the input low-light image, processing different frequency domain information on an image and feature level, and extracting appearance features of the image; designing a detail reservation module, and capturing detail information of different scales of the image by using a multi-scale convolution block to obtain detail features of the input image; a feature fusion module is designed, appearance features and detail features of the image are fused together through a soft gating fusion mechanism, and feature reconstruction is performed on a fusion result by using a multi-layer convolution block to obtain an enhancement result; and a self-adaptive loss function is designed, and a plurality of losses complement each other through shared information, so that the model is converged, and finally, the deep network of the optimal enhanced image is obtained. The model constructed by the invention can effectively solve the problems of poor image texture, large noise, color deviation and the like in the existing method, and better conforms to the enhancement effect of human perception.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer digital image processing, and in particular relates to a low-light image enhancement method based on a soft gating fusion mechanism and adaptive frequency domain perception. Background Art

[0002] In low-light conditions, due to insufficient light, low-light images often face problems such as low contrast, significant noise, color distortion, and reduced visibility. These problems not only affect the visual effect of the image, but also have a serious negative impact on the accuracy of downstream tasks, especially in key areas such as intelligent driving assistance systems, safety monitoring, photography art, and medical imaging. Therefore, low-light image enhancement technology has become an important and challenging research topic in the field of computer vision. Traditional methods include histogram equalization and retina theory (Retinex), but these methods have limitations in noise suppression and detail preservation. In recent years, with the development of deep learning technology, image enhancement methods based on convolutional neural networks (CNNs) have become a hot topic of research. These methods learn the multi-level features of images in pixel space and establish a complex mapping relationship from low-light images to normal light images, thereby effectively removing noise and significantly improving the quality of low-light images.

[0003] The existing deep learning image enhancement methods can be roughly divided into four categories, including: image enhancement methods based on convolutional neural networks, image enhancement methods based on CNN-Transformer, image enhancement methods based on models, and image enhancement methods based on fusion. However, most of the existing image enhancement technologies are enhanced in the spatial domain, which often leads to the loss of texture details, noise amplification, and color distortion while improving image quality. This is still a key problem that needs to be solved in the field of image enhancement. Summary of the invention

[0004] The purpose of the present invention is to provide a low-light image enhancement method based on a soft-gated fusion mechanism and adaptive frequency domain perception, which solves the problems of loss of texture details, noise amplification and color distortion often associated with existing image enhancement technologies.

[0005] The technical solution adopted by the present invention is: a low-light image enhancement method based on a soft-gated fusion mechanism and adaptive frequency domain perception, which adopts a frequency domain perception and detail retention module based on a diffusion model to effectively capture and enhance the key features of the image, deeply analyzes the image spectrum characteristics under different lighting conditions through wavelet transform, adaptively processes frequency domain information, and realizes preliminary enhancement of the image. At the same time, the detail retention module ensures that key information is not lost; in addition, the introduction of the soft-gated fusion mechanism allows the algorithm to significantly improve the visual quality of the image in a low-light environment while maintaining the natural texture of the image, thereby realizing the optimization and enhancement of the image; the specific operation steps are as follows:

[0006] Step 1: In the frequency domain space, a frequency domain perception module FDAM based on a diffusion model is designed. The frequency domain perception module effectively perceives and processes the high-frequency and low-frequency information of the image at the image and feature levels by performing frequency domain analysis on the input low-light images under different lighting conditions, thereby extracting the appearance features of the input image.

[0007] Step 2: In the pixel space, a detail preservation module DPM is designed. The detail preservation module DPM uses a multi-scale convolution block to capture detail information of different scales in the image, ensuring that the detail information is retained and enhanced to the greatest extent during the feature extraction process, thereby obtaining the detail features of the input image.

[0008] Step 3, designing a feature fusion module FFM, which complementarily fuses the appearance features and detail features of the image through a soft gated fusion mechanism, and then uses a multi-layer convolution block to reconstruct the features of the fusion result to obtain an enhanced result;

[0009] Step 4: Design a weight-adaptive loss function. Multiple losses share information and complement each other to improve each other’s performance, ultimately obtaining a deep network that outputs the optimal enhanced image.

[0010] Step 5, training the neural network built based on the soft gating fusion mechanism and the low-light image enhancement method based on the adaptive frequency domain perception for 200 epochs, verifying the training results, and saving the neural network model;

[0011] Step 6, test the neural network model saved in step 5 and output the enhanced image.

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

[0013] In step 1, the frequency domain perception module FDAM based on the diffusion model enhances the visual features of the low-light scene image through frequency domain analysis and processing; the frequency domain perception module FDAM includes three parts: frequency domain analysis block, frequency domain perception block and frequency domain parsing block; the frequency domain analysis block uses the existing discrete wavelet transform to perform frequency domain analysis on the input low-light scene image and converts the spatial domain features into frequency domain features; after obtaining the frequency domain features, the frequency domain perception block further perceives and enhances these frequency domain features, thereby highlighting the key visual features in the image and suppressing noise interference factors; finally, the frequency domain parsing block remaps the frequency domain features to the spatial features through the inverse discrete wavelet transform, thereby obtaining the final appearance features F a .

[0014] Step 1 The specific steps are as follows:

[0015] Step 1.1, firstly, the paired normal light images I in the low light scene n and low light images I l The input is sent to the frequency domain analysis block, and the features in the image domain are decomposed into different frequency components using a two-dimensional discrete wavelet transform, thereby extracting the frequency domain information W of the normal light image and the low light image. Ni and W Li , as shown in formula (1):

[0016]

[0017] Among them, W Ni and W Li They represent the frequency domain information of normal light and low light images respectively. L is the low frequency information of the image, which represents the downsampled version of the original image. H is the high frequency information detail information of the image, including the horizontal, vertical and diagonal edge information of the image.

[0018] Step 1.2: In the frequency domain perception block, the frequency domain information W of the normal light image is converted to Ni After the forward diffusion process q of stage T, a pure noise image W is obtained. T , as shown in formula (2); then W T and the frequency domain information W of the low-light image Li Perform a connection operation, such as contact(W T ,W Li ) to obtain the input feature W of the perception network θ input , as shown in formula (3):

[0019]

[0020] W input =concat(W T ,W Li )(3)

[0021] in, ε follows the standard normal distribution N(0,1); then the obtained joint input feature W input Input to the perception network θ, which consists of an encoder and a decoder and is used to perceive the frequency domain information of low-light scenes. In the encoding stage, the feature processing formula of each layer is shown in formula (4):

[0022] F out_i_h ,F out_i_l =DWT(Conv(Res(F in_i ,t))),(4)

[0023] F out_i_l =Conv(Res(F out_i_l ))+DWT(Conv(F in_i ))

[0024] where i∈{1,2,3,4}, F in_i and F out Represent the input features and output features of each layer respectively. Specifically, the input feature F in_i After residual processing and convolution mapping, the time t is decomposed into low-frequency information F by discrete wavelet transform. out_i_l and high frequency information F out_i_h , then, the low-frequency information is further processed to extract deeper features, thereby obtaining the final output feature F of the i-th layer out_i_l and F out_i_h ; F out_i_l The next layer module passed to the encoder continues to extract deep features, F out_i_l It is directly passed to the corresponding upsampling layer of the decoder through a skip connection to preserve the details of the image;

[0025] In the decoding stage, in order to enhance the image features and use high-frequency information to enhance the details of the image, the low-frequency features F generated by the downsampling block are out_i_l and high frequency features F out_i_h Combined, these features are mapped back to the original space using the inverse discrete wavelet transform (IWT) to generate upsampled feature maps. The processing flow of each layer is shown in formula (5):

[0026]

[0027] The core part of the perception network θ is the encoder and decoder, which are constructed through four layers of frequency domain perception up and down sampling modules, as shown in formula (4) and formula (5); the perception network finally constructed can accurately predict the mean and variance of the added noise at the tth moment of the diffusion process, as shown in formula (6):

[0028]

[0029] Among them, ε θ (.) refers to the output of the perception network θ. After being processed by formula (6), the corresponding mean and variance are obtained. Then, through the sampling process of the T stage, the enhanced frequency domain information W is obtained. en , the formula is as follows:

[0030]

[0031] Step 1.3: In order to facilitate effective fusion processing in subsequent steps, the frequency domain information W obtained in step 1.2 is transformed using the inverse wavelet transform technique IWT. en Perform feature conversion, which converts the frequency domain feature W en Remap back to pixel space to obtain appearance feature F a , the specific conversion process is shown in formula (8):

[0032] F a =IWT(W en ) (8)

[0033] Where IWT(·) represents the inverse wavelet transform, which generates the appearance feature F a .

[0034] Step 2 is as follows:

[0035] The goal of the detail preservation module DPM is to supplement the appearance features obtained in step 1 with details during the fusion phase to ensure that the detail information is preserved, as follows:

[0036] Step 2.1: Input low-light image I l Normalization is performed to adjust the image pixel values ​​to a uniform range, thereby balancing the brightness and contrast of the image. In addition, the depth and width of the network are increased through deep convolution operations. This process captures more complex image patterns and textures through multi-layer convolution and nonlinear transformations, as shown in formula (9):

[0037] F m =Conv(LayerNorm(I l )) m (9)

[0038] Among them, m is the number of layers of the convolutional block, I l represents the input low-light image, F m Represents the multi-scale features extracted by deep convolution and normalization operations;

[0039] Step 2.2, a gating block SG is proposed to adjust the relationship between input features, as shown in formula (10):

[0040] F out =f1*g+f2*(1-g) (10)

[0041] From the above formula, we can see that the gating block SG divides the obtained multi-scale features into two parts in the channel dimension, namely f1 and f2; then a gating coefficient is obtained through convolution operation and sigmoid (·) activation function. The gating coefficient g is used to process f1 and f2 to further improve the model's ability to express image features, thereby obtaining the output feature F of SG. out ;

[0042] In addition, in order to enable the model to automatically learn and emphasize more important features in image enhancement while suppressing unimportant features, the multi-scale feature F extracted in step 2.1 is m In the processing, it is proposed to use the spatial channel attention mechanism SCA to dynamically adjust the weights of features of different scales;

[0043] Finally, the output F of the gated block SG is multiplied by pixel level. out The output of spatial channel attention SCA (F m ) to obtain the deep feature F Y , as shown in formula (11):

[0044] F Y =F out *SCA(F m ) (11);

[0045] Among them, F Y is obtained by combining the output F of the gating block SG out The feature representation obtained by the output of the spatial channel attention SCA integrates the salient features of the image at different spatial locations and channels;

[0046] Step 2.3, then, after a series of normalization, convolution, activation and Dropout operations, the deep representation F obtained in step 2.2 is Y Processing is performed to ensure that the detailed information of the image is retained and enhanced to the greatest extent in the feature extraction stage, while enhancing the feature representation ability and suppressing overfitting; finally, the input image and the enhanced features are effectively combined through residual connection and two learnable parameters to obtain the detailed features F of the input low-light image d .

[0047] Step 3 is as follows:

[0048] Design a feature fusion module FFM to combine the appearance features F a and detail features F d The complementary features are fused together, and then the convolution block is used to reconstruct the features of the fusion result to obtain the enhanced result F en , specifically divided into two steps:

[0049] Step 3.1: Through the soft gating mechanism, the appearance features F extracted in step 1 are respectively a And the detail feature F extracted in step 2 d To perform interactive processing, firstly, the appearance feature F a and detail features F d Perform the splicing operation of the channel dimension, and then calculate the soft gate G of the appearance features through a convolution layer and a Sigmoid activation function. a and soft gating of detail features G d , the process is shown in formula (12):

[0050]

[0051] Among them, g(·) and h(·) represent the convolutional layers, and δ(·) is the Sigmoid activation function;

[0052] Step 3.2, use the soft gate G obtained in step 3.1 a Perform appearance feature F a With detail feature F d Perceptual fusion of the generated appearance-perceived detail features F d ' and detail-perceived appearance features F a ', as shown in formula (13):

[0053]

[0054] Among them, α is a training parameter with an initial value of 0, ⊙ represents element-wise multiplication, represents element-wise addition; then, and generates F d 'Same, using soft gating G of detail features d Control detail feature F d Fusion to appearance feature F a , the process is shown in formula (14):

[0055]

[0056] Among them, β is a training parameter with an initial value of 0;

[0057] Finally, the fused appearance features and detail features are spliced ​​to obtain a preliminary fusion result F combine, then the fusion result F combine After passing through 5 convolutional layers in sequence, the enhanced result F is obtained. en , the enhancement result is expressed as the following formula (15):

[0058] F en =(C,W,H) (15)

[0059] Among them, C represents the number of channels of the image, W represents the width of the image, and H represents the height of the image.

[0060] Step 4 is as follows:

[0061] Step 4.1, the loss function consists of three parts: noise loss L noise , frequency domain loss L frequency and content loss L content ; First, through L noise The difference between the perceptual network and the randomly generated target noise is calculated to constrain the prediction of the noise; secondly, L frequency The result of constrained frequency domain perception module FDAM; finally, using L content The results of the constrained frequency domain perception module FDAM, the detail preservation module DPM and the fusion module FFM are used to measure the similarity of the image content and structure of the output image. The loss functions of these three parts are optimized and a weight-adaptive loss function is designed, as shown in formula (16):

[0062] L sum =αL noise +βL frequency +λL content +b (16)

[0063] Among them, α, β and λ are the weights of each loss function respectively, and b is the bias of the loss function.

[0064] Step 4.2, Noise loss L noise The noise ε predicted by the perception network θ is calculated using the mean square error L2 θ The error information between the random noise I is obtained by minimizing L noise To improve the image quality, the noise loss L noise The calculation of is shown in formula (17):

[0065]

[0066] Among them, μ θ and is the noise ε predicted by the perception network θ θ The mean and variance of , I is the random noise added in the forward process that follows the standard normal distribution;

[0067] Step 4.3, frequency domain loss L frequency By calculating the mean square error L frequency_l2 and the total variation error L frequency_tv To constrain the enhanced appearance features Fa, L of the frequency domain perception module frequency The calculation formula is shown in formula (18):

[0068]

[0069] Among them, φ and is the weight of the loss function;

[0070] L frequency_l2 is the result of wavelet transform after denoising en The wavelet transform result of the target image gt dwt The mean square error of , thus ensuring that the low-frequency components and high-frequency components after denoising are close to the target image, as shown in formula (19):

[0071] L frequency_l2 =||w en ,gt dwt ||2(19)

[0072] The target image refers to a normal light image corresponding to the input low light image;

[0073] L frequency_tv is the result of wavelet transform after denoising en The total variation error with the input image ensures that the model effectively reduces the noise in the image and avoids excessive blurring of the image, as shown in formula (20):

[0074]

[0075] Where R(i,j)=W en (i,j)-I l_dwt (i,j), represents the result of wavelet transform after denoising W en With the input image I l Error information between wavelet transform results;

[0076] Step 4.4, L content The loss consists of two parts. First, the appearance feature F is calculated. a Detailed features F d And the enhanced result F en With the target image I nThe absolute error L1 between them ensures that the enhanced image is close to the target image at the pixel level; then, the structural similarity index SSIM is calculated to measure the structural similarity between the enhanced result of each module and the target image, making the enhanced image more natural visually; L content The calculation formula is shown in formula (21):

[0077]

[0078] Step 4.5, L noise , L frequency and L content By sharing information through adaptive weights, complementing each other, and updating network parameters, the network gradually converges.

[0079] Step 5 is as follows:

[0080] Step 5.1, the neural network uses the ADAM optimizer, the initial learning rate is set to 0.0002, and the learning rate scheduling strategy adopts the multi-step decay method, the step size is set to 50, and the decay factor is 0.8, that is, every 50 training steps, the learning rate will decay to 80% of the original; the entire training process lasts for 200 epochs;

[0081] Step 5.2, save the neural network parameters trained in step 5.1, the current training epoch, the optimizer ADAM used, and the scheduler to ensure that the training process and results are fully recorded. Finally, a complete and optimal enhanced network model is trained.

[0082] The beneficial effects of the present invention are:

[0083] The present invention is a low-light image enhancement method based on a soft gating fusion mechanism and adaptive frequency domain perception. It significantly improves the visual quality of low-light images by combining innovative methods in the frequency domain and the spatial domain. The core is to adopt a frequency domain perception and detail retention module based on a diffusion model to effectively capture and enhance the key features of the image. Through wavelet transform, the image spectrum characteristics under different lighting conditions are deeply analyzed, and the frequency domain information is adaptively processed to achieve preliminary enhancement of low-light images and obtain appearance features. At the same time, the detail retention module extracts detail features to ensure that key information is not lost. In the fusion stage, the soft gating mechanism optimizes the interaction between appearance and detail features, which not only retains the natural texture of the image, but also significantly improves the visual performance in low-light environments. In addition, the introduction of a weight-adaptive loss function further optimizes the enhancement effect of the network and solves common problems such as texture detail loss, noise amplification, and color distortion in image enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1It is a schematic diagram of the overall structure of the low-light image enhancement method based on the soft gating fusion mechanism and adaptive frequency domain perception of the present invention;

[0085] Figure 2 It is a structural schematic diagram of the frequency domain perception up and down sampling module in the low-light image enhancement method based on the soft gating fusion mechanism and adaptive frequency domain perception of the present invention;

[0086] Figure 3 It is a structural schematic diagram of a detail preservation module in a low-light image enhancement method based on a soft-gating fusion mechanism and adaptive frequency domain perception according to the present invention;

[0087] Figure 4 It is a structural schematic diagram of a fusion module in a low-light image enhancement method based on a soft-gating fusion mechanism and adaptive frequency domain perception according to the present invention;

[0088] Figure 5 It is a schematic diagram of changes in the evaluation indicators Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) during the training process of the low-light image enhancement method based on the soft-gated fusion mechanism and adaptive frequency domain perception of the present invention;

[0089] Figure 6 It is a schematic diagram of the effect of the low-light image enhancement method based on the soft gating fusion mechanism and adaptive frequency domain perception in low-light scenes. DETAILED DESCRIPTION

[0090] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0091] Example 1

[0092] The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception of the present invention has the following process: Figure 1 As shown in the figure, the frequency domain perception module FDAM based on the diffusion model is used to perceive and process the different frequency domain information of the low-light image in the frequency domain to obtain a preliminary enhanced appearance representation. The detail preservation module DPM is used to extract feature representations of different scales to obtain detail representation, thereby avoiding information loss in the feature processing process. The feature fusion module uses a soft gating interaction method to fuse the appearance features and detail features together, which maximizes the retention of the image's structural representation and edge representation. The convolution block is used for image reconstruction, which can accurately describe the detail features of the input low-light image to achieve low-light image enhancement. Finally, the weight of the loss function is dynamically adjusted to constrain the network to achieve the optimal enhancement result.

[0093] Example 2

[0094] The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception is implemented in the following steps:

[0095] Step 1: In the frequency domain space, a frequency domain awareness module (FDAM) based on a diffusion model is designed. FDAM effectively perceives and processes the high-frequency and low-frequency information of the image at the image and feature levels by performing frequency domain analysis on the input low-light images under different lighting conditions, thereby extracting the appearance features of the input image. The module consists of three key parts: frequency domain analysis block, frequency domain perception block, and frequency domain parsing block.

[0096] Step 1 is as follows:

[0097] Step 1.1, such as Figure 1 As shown, firstly, the normal light images I in the low light scene are paired n and low light images I l Input to the frequency domain analysis block. In order to enhance the content expression of the image, improve the edge and texture information of the image, and speed up the image reasoning speed, it is proposed to use the existing two-dimensional discrete wavelet transform to decompose the features in the image domain into different frequency components, thereby extracting the frequency domain information W of the low-light image and the normal light image Ni and W Li , in order to enrich the model's perception ability in the frequency domain, as shown in formula (1):

[0098]

[0099] Among them, W Ni and W Li They represent the frequency domain information of normal light and low light images respectively. L is the low-frequency information of the image, which represents the downsampled version of the original image. H is the high-frequency information detail information of the image, including the horizontal, vertical and diagonal edge information of the image.

[0100] Step 1.2: In the frequency domain perception block, the frequency domain information W of the normal light image is converted to Ni After the forward diffusion process q of stage T, a pure noise image W is obtained. T , as shown in formula (2); then W T and the frequency domain information W of the low-light image Li Perform a connection operation, such as contact(W T ,W Li ) to obtain the input feature W of the perception network θ input , as shown in formula (3).

[0101]

[0102] W input =concat(w T ,W Li )(3)

[0103] in, ε follows the standard normal distribution N(0,1); then the obtained joint input feature W input Input to the perception network θ, θ consists of an encoder and a decoder. In order to improve the model's ability to perceive frequency domain information at the feature level, that is, to improve the efficiency of the model while retaining high-frequency details, a new θ network is designed to perceive the frequency domain information of low-light scenes. Specifically, Figure 2 As shown in Figure 1, a frequency-domain-aware up- and down-sampling module and a skip residual block are proposed to process the input features. In the encoding stage, the feature processing formula of each layer is shown in formula (4):

[0104]

[0105] where i∈{1,2,3,4}, F in_i and F out Represent the input features and output features of each layer respectively. Specifically, the input feature F in_i After residual processing and convolution mapping, the time t is decomposed into low-frequency information F by discrete wavelet transform. out_i_l and high frequency information F out_i_h Subsequently, the low-frequency information is further processed to extract deeper features, thereby obtaining the final output feature F of the i-th layer out_i_l and F out_i_h . F out_i_l The next layer module passed to the encoder continues to extract deep features, F out_i_l It is directly passed to the corresponding upsampling layer of the decoder through a skip connection to preserve the details of the image.

[0106] In the decoding stage, in order to enhance the image features and use high-frequency information to enhance the details of the image, the low-frequency features F generated by the downsampling block are out_i_l and high frequency features F out_i_h Combined. Use the inverse discrete wavelet transform (IWT) to map these features back to the original space to generate upsampled feature maps. The processing flow of each layer is shown in formula (5):

[0107]

[0108] In summary, the core parts of the perception network θ: encoder and decoder, are constructed through a carefully designed four-layer frequency domain perception up and down sampling module, as shown in formula (4) and formula (5). This structure not only optimizes the data processing capability of the network, but also significantly improves its performance in image processing tasks. The perception network finally constructed can accurately predict the mean and variance of the added noise at the tth moment of the diffusion process, as shown in formula 6:

[0109]

[0110] where ε θ (.) refers to the output of the perception network θ. After being processed by formula (6), the corresponding mean and variance are obtained. Then, through the sampling process of the T stage, the enhanced frequency domain information W is obtained. en , the formula is as follows,

[0111]

[0112] Step 1.3: In order to facilitate effective fusion processing in subsequent steps, the frequency domain information W obtained in step 1.2 is transformed using the inverse wavelet transform technique IWT. en Perform feature conversion. This process converts the frequency domain feature W en Remap back to pixel space to obtain appearance feature F a , the specific conversion process is shown in formula (8):

[0113] F a =IWT(W en ) (8)

[0114] Where IWT(·) represents the inverse wavelet transform, which generates the appearance feature F a .

[0115] Step 2: In the pixel space, a detail preservation module (DPM) is designed. DPM uses multi-scale convolution blocks to capture detail information of different scales in the image, ensuring that the detail information is retained and enhanced to the greatest extent during the feature extraction process, thereby obtaining the detail features of the input image.

[0116] Step 2 is as follows:

[0117] Step 2.1, such as Figure 3 As shown in the detail preservation module DPM processing, in order to eliminate the brightness difference caused by illumination, the input low-light image I l Normalization (LayerNorm) is performed to adjust the image pixel values ​​to a uniform range, thereby balancing the brightness and contrast of the image. In addition, the depth and width of the network are increased through deep convolution operations. This process can capture more complex image patterns and textures through multi-layer convolution and nonlinear transformation, as shown in formula (9):

[0118] F m =Conv(LayerNorm(I l )) m (9)

[0119] Where m is the number of layers of the convolutional block, I l represents the input low-light image, F m Represents the multi-scale features extracted by deep convolution and normalization operations.

[0120] Step 2.2: In order to enhance the correlation between the multi-scale features of the input, a simple gating block (SG) is first proposed to adjust the relationship between the input features, as shown in formula (10):

[0121] F out =f1*g+f2*(1-g) (10)

[0122] From the above formula, we can see that the gating block SG divides the obtained multi-scale features into two parts in the channel dimension, namely f1 and f2. Then, through the convolution operation and the sigmoid (·) activation function, a gating coefficient is obtained. The gating coefficient g is used to process f1 and f2 to further improve the model's ability to express image features, thereby obtaining the output feature F of SG. out .

[0123] In addition, in order to enable the model to automatically learn and emphasize more important features in image enhancement while suppressing unimportant features, the multi-scale feature F extracted in step 2.1 is m In the processing of image enhancement, it is proposed to use the spatial channel attention mechanism (SCA) to dynamically adjust the weights of features of different scales, so that the network pays more attention to the most useful features for image enhancement tasks, better preserves image details and suppresses noise.

[0124] Finally, the output F of the gated block SG is multiplied by pixel level. out The output of spatial channel attention SCA (F m ) to obtain the deep feature F Y , for subsequent processing, as shown in formula (11):

[0125] F Y =F out *SCA(F m )(11);

[0126] The SCA in the formula can highlight important areas and features based on the spatial and channel information of the image. The output F of the gating network out It can dynamically adjust the weights of different features to further enhance important features and suppress irrelevant features. The above uses spatial channel attention and a simple gating mechanism to enhance the features of key areas in the image and improve the contrast of the image. Yis obtained by combining the output F of the gating block SG out The feature representation obtained by the output of the spatial channel attention SCA integrates the significant features of the image at different spatial positions and channels, making it more suitable for subsequent image processing tasks.

[0127] Step 2.3, then, after a series of normalization, convolution, activation and Dropout operations, the deep representation F obtained in step 2.2 is Y Processing is performed to ensure that the detailed information of the image is retained and enhanced to the greatest extent in the feature extraction stage, while enhancing the feature representation ability and suppressing overfitting; finally, the input image and the enhanced features are effectively combined through residual connection and two learnable parameters to obtain the detailed features F of the input low-light image d .

[0128] Step 3, such as Figure 4 As shown in the figure, a feature fusion module (FFM) is designed. FFM complementarily fuses the appearance features and detail features of the image through a soft gating fusion mechanism, and then uses a multi-layer convolution block to reconstruct the features of the fusion result to obtain an enhanced result.

[0129] Step 3 is as follows:

[0130] Step 3.1: Through the soft gating mechanism, the appearance features F extracted in step 1 are respectively a And the detail feature F extracted in step 2 d Specifically, firstly, the appearance feature F a and detail features F d Perform the splicing operation of the channel dimension, and then calculate the soft gate G of the appearance features through a convolution layer and a Sigmoid activation function. a and soft gating of detail features G d , the process is shown in formula (12):

[0131]

[0132] Here g(·) and h(·) represent the convolutional layers, and δ(·) is the Sigmoid activation function. This process ensures that the fused features contain both global appearance information and rich detail features, improving the overall quality and detail accuracy of the final generated image.

[0133] Step 3.2, use the soft gating obtained in step 3.1 to perform perceptual fusion of appearance features and detail features (such as Figure 4 As shown), generate appearance-aware detail features F d ' and detail-perceived appearance features Fa '. First use soft gating G a Control appearance features F a Fusion into detail features F d , the process is shown in formula (13):

[0134]

[0135] Where α is a training parameter with an initial value of 0, ⊙ represents element-wise multiplication, represents element-wise addition. Then, with F d 'Same, using soft gating G of detail features d Control detail feature F d Fusion to appearance feature F a , the process is shown in formula (14):

[0136]

[0137] Where β is a training parameter with an initial value of 0.

[0138] Finally, the fused appearance features and detail features are spliced ​​to obtain a preliminary fusion result F combine , then the fusion result F combine After passing through 5 convolutional layers in sequence, the enhanced result F is obtained. en , the enhancement result is expressed as the following formula (15):

[0139] F en =(C,W,H) (15)

[0140] Among them, C represents the number of channels of the image, W represents the width of the image, and H represents the height of the image.

[0141] Step 4: Design a weight-adaptive loss function. Multiple losses can improve each other’s performance by sharing information and complementing each other, and finally obtain a deep network that outputs the optimal enhanced image.

[0142] Step 4 is as follows:

[0143] Step 4.1, such as Figure 1 As shown, the loss function consists of three parts: noise loss L noise , frequency domain loss L frequency and content loss L content ; First, through L noise The difference between the perceptual network and the randomly generated target noise is calculated to constrain the prediction of the noise. frequency The result of the constrained frequency domain aware module (FDAM). Finally, using L contentThe results of the frequency domain perception module (FDAM), detail preservation module (DPM) and fusion module (FFM) are constrained to measure the similarity of the image content and structure of the output image. The three loss functions are optimized and a weight-adaptive loss function is designed, as shown in formula (16):

[0144] L sum =αL noise +βL frequency +λL content +b (16)

[0145] Among them, α, β and λ are the weights of each loss function respectively, and b is the bias of the loss function.

[0146] Step 4.2, Noise loss L noise The noise ε predicted by the perception network θ is calculated using the mean square error L2 θ The error information between the random noise I is obtained by minimizing L noise To improve the image quality. Noise loss L noise The calculation of is shown in formula (17):

[0147]

[0148] Among them, μ θ and is the noise ε predicted by the perception network θ θ The mean and variance of , I is the random noise added in the forward process that follows the standard normal distribution.

[0149] Step 4.3, frequency domain loss L frequency By calculating the mean square error L frequency_l2 and the total variation error L frequency_tv To constrain the enhanced appearance features Fa, L of the frequency domain perception module frequency The calculation formula is shown in formula (18):

[0150]

[0151] Among them, φ and is the weight of the loss function;

[0152] L frequency_l2 Calculate the result of wavelet transform after denoising W en The wavelet transform result of the target image gt dwt The mean square error of , thus ensuring that the low-frequency components and high-frequency components after denoising are close to the target image, as shown in formula (19):

[0153] L frequency_l2 =||w en ,gtdwt ||2(19)

[0154] The target image refers to a normal light image corresponding to the input low light image;

[0155] L frequency_tv Calculate the result of wavelet transform after denoising W en The total variation error of the input image is calculated to ensure that the model effectively reduces the noise in the image and avoids excessive blurring of the image, as shown in formula (20):

[0156]

[0157] Where R(i,j)=W en (i,j)-I l_dwt (i,j), represents the result of wavelet transform after denoising W en With the input image I l Error information between wavelet transform results.

[0158] Step 4.4, in order to measure the similarity between the enhanced image and the target image at the pixel level and structure level, we use the content loss L in the FDAM, DPM and FFM modules content To know that the network generates high-quality images. Specifically, L content The loss consists of two parts. First, the appearance feature F is calculated. a Detailed features F d And the enhanced result F en With the target image I n The absolute error L1 between them ensures that the enhanced image is close to the target image at the pixel level; then, the structural similarity index SSIM is calculated to measure the structural similarity between the enhanced result of each module and the target image, making the enhanced image more natural visually; L content The calculation formula is shown in formula (21):

[0159]

[0160] Step 4.5, L sum Dynamic learning loss weight, L noise , L frequency and L content By sharing information through adaptive weights, complementing each other, and updating network parameters, the network gradually converges.

[0161] Step 5, training the neural network built based on the soft gating fusion mechanism and the low-light image enhancement method based on the adaptive frequency domain perception for 200 epochs, verifying the training results, and saving the neural network model;

[0162] Step 5 is as follows:

[0163] Step 5.1, select a GPU with 30GB video memory for training, the network uses the ADAM optimizer, and the initial learning rate is set to 0.0002. The learning rate scheduling strategy uses a multi-step decay method, with a step size of 50 and a decay factor of 0.8, that is, every 50 training steps, the learning rate will decay to 80% of the original. The entire training process lasts 200 epochs.

[0164] Step 5.2: Save the neural network parameters trained in step 5.1, the current training epoch, the optimizer ADAM used, and the scheduler to ensure that the training process and results are fully recorded. After each round of training, save the PSNR and SSIM values, and draw the curves of the structural similarity SSIM and peak signal-to-noise ratio PSNR for 200 rounds. Figure 5 As shown in the figure, the indicators fluctuated greatly before the 100th round, but were generally on an upward trend. After the 170th round, they tended to stabilize, with the final PSNR value of 25.0090 and the SSIM value of 0.8614. Finally, a complete and optimal enhanced network model was trained to facilitate subsequent verification, testing, and practical applications.

[0165] Step 6, test the neural network model saved in step 5 and output the enhanced image.

[0166] Step 6 is as follows:

[0167] Load the network model trained in step 5, input the test set into the trained network model, and then save the test results to obtain the enhanced image.

[0168] Example 3

[0169] like Figure 6 The figure shows the effect diagram of the present invention in a low-light scene. Each row in the figure represents images in the same data set. The first column of images is the input low-light image, the second column of images is the result of restoration using the method of the present invention, and the third column of images is the normal light image corresponding to the first column of images.

[0170] It can be seen that the low-light image enhancement method based on the soft gating fusion mechanism and adaptive frequency domain perception of the present invention adopts the frequency domain perception and detail retention module based on the diffusion model to effectively capture and enhance the key features of the image. The network can deeply analyze the image spectrum characteristics under different lighting conditions, adaptively process the frequency domain information, achieve preliminary enhancement of low-light images, and obtain appearance features. At the same time, the detail retention module extracts detail features to ensure that key information is not lost. In the fusion stage, the soft gating mechanism optimizes the interaction between appearance and detail features, which not only retains the natural texture of the image, but also significantly improves the visual performance in low-light environments. In addition, the introduction of a weight-adaptive loss function further optimizes the enhancement effect of the network, solves common problems such as texture detail loss, noise amplification, and color distortion in image enhancement, and has certain practical significance.

Claims

1. A low-light image enhancement method based on soft-gating fusion mechanism and adaptive frequency domain perception, characterized in that: The frequency domain perception and detail preservation module based on the diffusion model is adopted to effectively capture and enhance the key features of the image. The wavelet transform is used to deeply analyze the image spectrum characteristics under different lighting conditions, and the frequency domain information is adaptively processed to achieve preliminary image enhancement. At the same time, the detail preservation module ensures that key information is not lost. In addition, the introduction of the soft gating fusion mechanism allows the algorithm to significantly improve the visual quality of the image in low-light environments while maintaining the natural texture of the image, thereby achieving optimized image enhancement.

2. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 1 is characterized in that: The specific steps are as follows: Step 1: In the frequency domain space, a frequency domain perception module FDAM based on a diffusion model is designed. The frequency domain perception module effectively perceives and processes the high-frequency and low-frequency information of the image at the image and feature levels by performing frequency domain analysis on the input low-light images under different lighting conditions, thereby extracting the appearance features of the input image. Step 2: In the pixel space, a detail preservation module DPM is designed. The detail preservation module DPM uses a multi-scale convolution block to capture detail information of different scales in the image, ensuring that the detail information is retained and enhanced to the greatest extent during the feature extraction process, thereby obtaining the detail features of the input image. Step 3, designing a feature fusion module FFM, which complementarily fuses the appearance features and detail features of the image through a soft gated fusion mechanism, and then uses a multi-layer convolution block to reconstruct the features of the fusion result to obtain an enhanced result; Step 4: Design a weight-adaptive loss function. Multiple losses share information and complement each other to improve each other’s performance, ultimately obtaining a deep network that outputs the optimal enhanced image. Step 5, training the neural network built based on the soft gating fusion mechanism and the low-light image enhancement method based on the adaptive frequency domain perception for 200 epochs, verifying the training results, and saving the neural network model; Step 6, test the neural network model saved in step 5 and output the enhanced image.

3. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 2 is characterized in that: In step 1, the frequency domain perception module FDAM based on the diffusion model enhances the visual features of the low-light scene image through frequency domain analysis and processing; the frequency domain perception module FDAM includes three parts: frequency domain analysis block, frequency domain perception block and frequency domain parsing block; the frequency domain analysis block uses the existing discrete wavelet transform to perform frequency domain analysis on the input low-light scene image and converts the spatial domain features into frequency domain features; After obtaining the frequency domain features, the frequency domain perception block further perceives and enhances these frequency domain features, thereby highlighting the key visual features in the image and suppressing noise interference factors; finally, the frequency domain analysis block remaps the frequency domain features to spatial features through inverse discrete wavelet transform, thereby obtaining the final appearance feature F a .

4. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 3 is characterized in that: Step 1 The specific steps are as follows: Step 1.1, firstly, the paired normal light images I in the low light scene n and low light images I l The input is sent to the frequency domain analysis block, and the features in the image domain are decomposed into different frequency components using a two-dimensional discrete wavelet transform, thereby extracting the frequency domain information W of the normal light image and the low light image. Ni and W Li , as shown in formula (1): Among them, W Ni and W Li They represent the frequency domain information of normal light and low light images respectively. L is the low frequency information of the image, which represents the downsampled version of the original image. H is the high frequency information detail information of the image, including the horizontal, vertical and diagonal edge information of the image. Step 1.2: In the frequency domain perception block, the frequency domain information W of the normal light image is converted to Ni After the forward diffusion process q of stage T, a pure noise image W is obtained. T , as shown in formula (2); then W T and the frequency domain information W of the low-light image Li Perform a connection operation, such as contact(W T ,W Li ) to obtain the input feature W of the perception network θ input , as shown in formula (3): W input =concat(W T ,W Li ) (3) in, ε follows the standard normal distribution N(0,1); then the obtained joint input feature W input Input to the perception network θ, which consists of an encoder and a decoder and is used to perceive the frequency domain information of low-light scenes. In the encoding stage, the feature processing formula of each layer is shown in formula (4): where i∈{1,2,3,4}, F in_i and F out Represent the input features and output features of each layer respectively. Specifically, the input feature F in_i After residual processing and convolution mapping, the time t is decomposed into low-frequency information F by discrete wavelet transform. out_i_l and high frequency information F out_i_h , then, the low-frequency information is further processed to extract deeper features, thereby obtaining the final output feature F of the i-th layer out_i_l and F out_i_h ; F out_i_l The next layer module passed to the encoder continues to extract deep features, F out_i_l It is directly passed to the corresponding upsampling layer of the decoder through a skip connection to preserve the details of the image; In the decoding stage, in order to enhance the image features and use high-frequency information to enhance the details of the image, the low-frequency features F generated by the downsampling block are out_i_l and high frequency features F out_i_h Combined, these features are mapped back to the original space using the inverse discrete wavelet transform (IWT) to generate upsampled feature maps. The processing flow of each layer is shown in formula (5): The core part of the perception network θ is the encoder and decoder, which are constructed through four layers of frequency domain perception up and down sampling modules, as shown in formula (4) and formula (5); the perception network finally constructed can accurately predict the mean and variance of the added noise at the tth moment of the diffusion process, as shown in formula (6): Among them, ε θ (.) refers to the output of the perception network θ. After being processed by formula (6), the corresponding mean and variance are obtained. Then, through the sampling process of the T stage, the enhanced frequency domain information W is obtained. en , the formula is as follows: Step 1.3: In order to facilitate effective fusion processing in subsequent steps, the frequency domain information W obtained in step 1.2 is transformed using the inverse wavelet transform technique IWT. en Perform feature conversion, which converts the frequency domain feature W en Remap back to pixel space to obtain appearance feature F a , the specific conversion process is shown in formula (8): F a =IWT(W en )(8) Where IWT(·) represents the inverse wavelet transform, which generates the appearance feature F a .

5. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 2, characterized in that: Step 2 is as follows: The goal of the detail preservation module DPM is to supplement the appearance features obtained in step 1 with details during the fusion phase to ensure that the detail information is preserved, as follows: Step 2.1: Input low-light image I l Normalization is performed to adjust the image pixel values ​​to a uniform range, thereby balancing the brightness and contrast of the image. In addition, the depth and width of the network are increased through deep convolution operations. This process captures more complex image patterns and textures through multi-layer convolution and nonlinear transformations, as shown in formula (9): F m =Conv(LayerNorm(I l )) m (9) Among them, m is the number of layers of the convolutional block, I l represents the input low-light image, F m Represents the multi-scale features extracted by deep convolution and normalization operations; Step 2.2, a gating block SG is proposed to adjust the relationship between input features, as shown in formula (10): F out =f1*g+f2*(1-g) (10) From the above formula, we can see that the gating block SG divides the obtained multi-scale features into two parts in the channel dimension, namely f1 and f2; then a gating coefficient is obtained through convolution operation and sigmoid (·) activation function. The gating coefficient g is used to process f1 and f2 to further improve the model's ability to express image features, thereby obtaining the output feature F of SG. out; In the multi-scale feature F extracted in step 2.1 m In the processing, it is proposed to use the spatial channel attention mechanism SCA to dynamically adjust the weights of features of different scales; Finally, the output F of the gated block SG is multiplied by pixel level. out The output of spatial channel attention SCA (F m ) to obtain the deep feature F Y , as shown in formula (11): F Y =F out *SCA(F m ) (11); Among them, F Y is obtained by combining the output F of the gating block SG out The feature representation obtained by the output of the spatial channel attention SCA integrates the salient features of the image at different spatial locations and channels; Step 2.3, then, after a series of normalization, convolution, activation and Dropout operations, the deep representation F obtained in step 2.2 is Y Processing is performed to ensure that the detailed information of the image is retained and enhanced to the greatest extent in the feature extraction stage, while enhancing the feature representation ability and suppressing overfitting; finally, the input image and the enhanced features are effectively combined through residual connection and two learnable parameters to obtain the detailed features F of the input low-light image d .

6. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 2, characterized in that: Step 3 is as follows: Design a feature fusion module FFM to combine the appearance features F a and detail features F d The complementary features are fused together, and then the convolution block is used to reconstruct the features of the fusion result to obtain the enhanced result F en , specifically divided into two steps: Step 3.1: Through the soft gating mechanism, the appearance features F extracted in step 1 are respectively a And the detail feature F extracted in step 2 d To perform interactive processing, firstly, the appearance feature F a and detail features F d Perform the splicing operation of the channel dimension, and then calculate the soft gate G of the appearance features through a convolution layer and a Sigmoid activation function. a and soft gating of detail features G d , the process is shown in formula (12): Among them, g(·) and h(·) represent the convolutional layers, and δ(·) is the Sigmoid activation function; Step 3.2, use the soft gate G obtained in step 3.1 a Perform appearance feature F a With detail feature F d Perceptual fusion of the generated appearance-perceived detail features F' d and detail-perceived appearance features F' a , as shown in formula (13): Among them, α is a training parameter with an initial value of 0, ⊙ represents element-wise multiplication, represents element-wise addition; then, and generates F' d Similarly, the soft gating G using detail features d Control detail feature F d Fusion to appearance feature F a , the process is shown in formula (14): Among them, β is a training parameter with an initial value of 0; Finally, the fused appearance features and detail features are spliced ​​to obtain a preliminary fusion result F combine , then the fusion result F combine After passing through 5 convolutional layers in sequence, the enhanced result F is obtained. en , the enhancement result is expressed as the following formula (15): F en =(C,W,H) (15) Among them, C represents the number of channels of the image, W represents the width of the image, and H represents the height of the image.

7. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 2, characterized in that: Step 4 is as follows: Step 4.1, the loss function consists of three parts: noise loss L noise , frequency domain loss L frequency and content loss L content ; First, through L noise Compute the difference between the perceptual network and a randomly generated target noise, thereby constraining the prediction of the noise; Secondly, through L frequency The result of constrained frequency domain perception module FDAM; finally, using L content The results of the constrained frequency domain perception module FDAM, the detail preservation module DPM and the fusion module FFM are used to measure the similarity of the image content and structure of the output image. The loss functions of these three parts are optimized and a weight-adaptive loss function is designed, as shown in formula (16): L sum =αL noise +βL frequency +λL content +b (16) Among them, α, β and λ are the weights of each loss function, and b is the bias of the loss function; Step 4.2, Noise loss L noise The noise ε predicted by the perception network θ is calculated using the mean square error L2 θ The error information between the random noise I is obtained by minimizing L noise To improve the image quality, the noise loss L noise The calculation of is shown in formula (17): Among them, μ θ and is the noise ε predicted by the perception network θ θ The mean and variance of , I is the random noise added in the forward process that follows the standard normal distribution; Step 4.3, frequency domain loss L frequency By calculating the mean square error L frequency_l2 and the total variation error L frequency_tv To constrain the enhanced appearance features Fa, L of the frequency domain perception module frequency The calculation formula is shown in formula (18): Among them, φ and is the weight of the loss function; L frequency_l2 is the result of wavelet transform after denoising en The wavelet transform result of the target image gt dwt The mean square error of , thus ensuring that the low-frequency components and high-frequency components after denoising are close to the target image, as shown in formula (19): L frequency_l2 =L2(w en ,gt dwt )=||w en ,gt dwt ||2 (19) The target image refers to a normal light image corresponding to the input low light image; L frequency_tv is the result of wavelet transform after denoising en The total variation error with the input image ensures that the model effectively reduces the noise in the image and avoids excessive blurring of the image, as shown in formula (20): Where R(i,j)=W en (i,j)-I l_dwt (i,j), represents the result of wavelet transform after denoising W en With the input image I l Error information between wavelet transform results; Step 4.4, L content The loss consists of two parts. First, the appearance feature F is calculated. a Detailed features F d And the enhanced result F en With the target image I n The absolute error L1 between them ensures that the enhanced image is close to the target image at the pixel level; then, the structural similarity index SSIM is calculated to measure the structural similarity between the enhanced result of each module and the target image, making the enhanced image more natural visually; L content The calculation formula is shown in formula (21): Step 4.5, L noise , L frequency and L content By sharing information through adaptive weights, complementing each other, and updating network parameters, the network gradually converges.

8. The low-light image enhancement method based on soft gating fusion mechanism and adaptive frequency domain perception according to claim 2, characterized in that: Step 5 is as follows: Step 5.1, the neural network uses the ADAM optimizer, the initial learning rate is set to 0.0002, and the learning rate scheduling strategy adopts the multi-step decay method, the step size is set to 50, and the decay factor is 0.8, that is, every 50 training steps, the learning rate will decay to 80% of the original; the entire training process lasts for 200 epochs; Step 5.2, save the neural network parameters trained in step 5.1, the current training epoch, the optimizer ADAM used, and the scheduler to ensure that the training process and results are fully recorded. Finally, a complete and optimal enhanced network model is trained.

Citation Information

Cited By

  • Microorganism image enhancement processing method

    CN119477735A

  • Dark light field image compression method based on probability representation

    CN119854494A

  • Low-light image enhancement method based on combination of Retinex and wavelet transform

    CN119963465A

  • Image enhancement method and system based on detail sensitivity and noise suppression fusion

    CN120563362A

  • Image enhancement method and system based on detail sensitivity and noise suppression fusion

    CN120563362B