Image quality optimization method based on multi-branch collaboration

Abstract

The invention relates to image optimization, in particular to an image quality optimization method based on multi-branch collaboration, which comprises the following steps of: inputting a denoising feature into a local information enhancement branch, and enhancing local texture information of a dark region to obtain a local enhancement feature while keeping the stability of an original structure of a low-light image; inputting the de-noising features into a global information enhancement branch, and performing global color information modeling and compensation on the de-noising features for the difference of the low-light image in the aspects of overall brightness distribution and color consistency to obtain color features; denoising features are input into a texture enhancement branch, high-frequency structure information is intensified in a targeted mode through a feature enhancement mechanism combining explicit texture guidance and implicit attention modeling, so that lost detail edge and local texture information of a low-light image after denoising is restored, the definition and visual detail expressive force of a finally reconstructed image are improved, and the image quality is improved. Obtaining texture features; the method can overcome the defect that brightness enhancement, texture recovery and color correction cannot be realized at the same time.

Description

Technical Field

[0001] This invention relates to image optimization, and more specifically to an image quality optimization method based on multi-branch collaboration. Background Technology

[0002] In 2023, Xin Jin et al. proposed "DNF: Decouple and Feedback Network for Seeingin the Dark," a low-light image enhancement method that employs a single enhancement path. This makes it difficult to simultaneously address dark noise suppression, texture restoration, and color naturalness, resulting in enhanced images still exhibiting detail blurring and color cast. Under extreme low-light conditions, due to limited input information and insufficient network feature representation capabilities, the enhancement results show weak adaptability to various scene variations. While existing techniques can denoise in the RAW domain and restore color in the sRGB domain, they still primarily focus on the basic reconstruction process, lacking a clear post-processing mechanism to improve the perceptual quality of the final image. Furthermore, feedback information typically originates from intermediate feature layers, lacking high-level feedback control geared towards output image quality, leaving significant room for improvement in detail sharpness and color fidelity.

[0003] Therefore, there is an urgent need for an image quality optimization method with multi-branch collaborative enhancement capabilities and a scalable feedback mechanism to simultaneously achieve brightness enhancement, texture restoration, and color correction. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides an image quality optimization method based on multi-branch collaboration, which can effectively overcome the shortcomings of the existing technology that cannot simultaneously achieve brightness enhancement, texture restoration and color correction.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: The image quality optimization method based on multi-branch collaboration includes the following steps: S1. Perform light magnification processing on the input low-light image; S2. Input the enhanced low-light image into the denoising decoding branch to effectively suppress noise and preserve image structure information, thereby obtaining denoising features; S3. Input the denoising features into the local information enhancement branch. While maintaining the stability of the original structure of the low-light image, enhance the local texture information in the dark area and provide local enhancement features for subsequent feature fusion and image reconstruction. S4. Input the denoising features into the global information enhancement branch. Based on the differences in the overall brightness distribution and color consistency of the low-light image, perform global color information modeling and compensation on the denoising features to obtain color features. S5. Input the denoising features into the texture enhancement branch. By combining the feature enhancement mechanism of explicit texture guidance and implicit attention modeling, the high-frequency structural information is specifically enhanced to restore the detail edges and local texture information lost in the low-light image after denoising, thereby improving the clarity and visual detail of the final reconstructed image and obtaining texture features. S6. Input texture features and color features into the fusion module. Through multi-level attention modeling and convolution transformation, the local texture information and global color information are collaboratively modeled and effectively fused to obtain fusion features that effectively enhance local texture information while ensuring color consistency.

[0006] Preferably, S1 involves performing light amplification processing on the input low-light image, including: To address underexposure caused by low-light environments, the input low-light image I is enhanced using a linear transformation combined with adaptive brightness adjustment, resulting in the enhanced low-light image I. light : ; ; Where ratio is the exposure magnification factor. , These are the exposure times for normal light images and low light images, respectively.

[0007] Preferably, in S2, the enhanced low-light image is input into the denoising decoding branch to effectively suppress noise and preserve image structural information, resulting in denoising features, including: S21, Enhance the low-light image I light Input encoder to extract initial features F enc : ; S22, Regarding the initial feature F enc Denoising and decoding are performed to obtain the denoised features F. denoised .

[0008] Preferably, in S3, the denoising features are input into the local information enhancement branch. While maintaining the stability of the original structure of the low-light image, the local texture information in the dark areas is enhanced, providing local enhancement features for subsequent feature fusion and image reconstruction, including: S31. Considering the severe loss of texture information in dark areas, an adaptive convolutional kernel prediction mechanism is introduced to specifically enhance the details in high-noise and low-brightness areas, resulting in the basic convolutional feature F. con : ; Among them, the basic convolutional feature F con By analyzing the denoising feature F denoised Perform two 3x3 convolutions With nonlinear activation This is obtained after performing local context modeling and feature reconstruction; S32. Introduce a channel attention mechanism to adaptively model the importance of different channels, and apply this to the denoising feature F. denoised Global average pooling (GAP) is performed to obtain channel-level statistics, and channel weights are generated using a multilayer perceptron (MLP) and a sigmoid function to optimize the basic convolutional features. Channel-by-channel recalibration is performed to highlight channels that are more critical for restoring texture information in dark areas, while suppressing redundant and noisy channels, resulting in the channel recalibration feature F. ca : ; in, The Hadamard product represents element-wise multiplication. S33. Introduce a spatial attention mechanism by recalibrating the channel features F. ca Average pooling (AvgPool) and max pooling (MaxPool) are performed separately to aggregate spatial dimensionality information, and then a 7x7 convolution is applied. The spatial weight map is generated using the Sigmoid function, and the channel features F are recalibrated. ca Adaptive enhancement is performed in spatial location to strengthen the effective texture structure in high-noise and low-brightness areas and improve the discernibility of texture information in dark areas, resulting in spatial recalibration features F. sa : ; S34. Spatial recalibration feature F sa With denoising feature F denoised Perform residual fusion and then use 1*1 convolution. Channel compression and feature integration are performed to obtain the local enhanced feature F. local : .

[0009] Preferably, in S4, the denoising features are input into the global information enhancement branch. Addressing the differences in overall brightness distribution and color consistency in low-light images, global color information modeling and compensation are performed on the denoising features to obtain color features, including: S41. Denoising feature F is achieved by using multi-layer convolution and non-linear activation LeakyReLU. denoised Initial mapping is performed, and the denoised features F are processed through local receptive field convolution.denoised Smoothing and recoding are performed to extract globally enhanced intermediate features F that are beneficial for modeling overall brightness and color distribution. Re : ; S42, Based on globally enhanced intermediate features F Re The generated channel weights m(F) Re The weighted values ​​are then processed, and the weighted results are subjected to global average pooling (GAP) in the spatial dimension to obtain statistical features reflecting the overall brightness and color distribution. Finally, a 1x1 convolution is applied. The Sigmoid function generates global context compensation information Context(F) Re ): ; S43. Transfer global context compensation information Context(F) Re Globally enhanced intermediate features F Re and denoising features F denoised By adding elements one by one, we obtain the color feature F, which integrates the overall brightness distribution and color consistency information. color : .

[0010] Preferably, in S5, the denoising features are input into the texture enhancement branch. By combining explicit texture guidance and implicit attention modeling as a feature enhancement mechanism, high-frequency structural information is specifically enhanced to recover the detail edges and local texture information lost in the low-light image after denoising, thereby improving the clarity and visual detail representation of the final reconstructed image and obtaining texture features, including: S51, Regarding the denoising feature F denoised Calculate the Sobel gradient to further recover detailed edges and local texture information, and obtain the gradient magnitude G: ; in, , These are the Sobel gradient operators in the X and Y directions, respectively. It is a positive number that approaches 0, used to avoid numerical instability when the gradient is zero; Normalizing the gradient magnitude G eliminates the influence of gradient magnitude scale differences between different images or regions, resulting in texture gradient features that are more suitable as structural prior information for subsequent attention calculations while maintaining the relative texture distribution. : ; Where max(G) is the maximum value in the gradient magnitude G; S52, Texture gradient features With denoising feature F denoised The system concatenates data along the channel dimension to construct a joint feature representation containing local texture and semantic information. Pointwise convolution (PWConv) is then used for channel mapping and dimensionality compression. Finally, depthwise separable convolution (DWConv) is employed to extract local spatial relationships, generating a query vector Q, a key vector K, and a value vector V for the attention mechanism. ; in, This indicates that the texture gradient feature With denoising feature F denoised The components are stitched together along the channel dimension. S53. Calculate the attention weight matrix Attn along the channel dimension to measure the importance of different channels in the texture enhancement task, and obtain the texture-aware enhancement feature F by weighted summation. attn This allows the network to adaptively focus on the channels most critical for texture restoration, suppressing ineffective or noise-dominated effects. ; ; in, This is a temperature coefficient used to adjust the smoothness of attention distribution. The Softmax function is used to ensure that the attention weights have interpretable probability distribution characteristics; S54. Integrate gradient magnitude G with texture-aware enhancement feature F attn By fusing explicit texture guidance and implicit attention modeling, the system effectively enhances realistic edges and texture structures while avoiding over-sharpening, resulting in the fused enhanced feature F. enh : ; in, This indicates a convolution mapping of the gradient magnitude G. These are the weighting coefficients. ; S55. A feedforward network is introduced at the end of the texture enhancement branch, and then connected to the denoising feature F through residual connections. denoised The features are fused to further enhance their representational power and stabilize the training process, resulting in texture features F. texture : .

[0011] Preferably, in step S6, texture features and color features are input into the fusion module. Through multi-level attention modeling and convolutional transformation, local texture information and global color information are collaboratively modeled and effectively fused to obtain fused features that effectively enhance local texture information while ensuring color consistency, including: S61, Considering texture features F texture With color feature F color The importance of different channel dimensions varies significantly. First, channel attention modeling is performed on the two types of features, and then for the texture feature F... texture Global average pooling (GAP) is performed to obtain channel-level statistics, and then channel convolution (Conv) is applied. c The Sigmoid function generates channel weights for the texture feature F. texture Weighted modulation is performed to obtain the texture attention feature F' texture : ; For color feature F color Using texture feature F texture Using the same channel attention modeling method, we obtain the color attention feature F'. color : ; Through the above processing, the network is able to adaptively emphasize channels that are more critical to texture restoration or color consistency; S62, Apply texture attention feature F' texture With color attention feature F' color The data is concatenated along the channel dimension, and then channel recombination and dimensionality compression are performed using pointwise convolution PWConv to obtain the intermediate feature F. pw : ; S63. To further model local spatial relationships, a depthwise separable convolution (DWConv) is introduced to model intermediate features F. pw The transformation is performed, splitting the output features into two parts along the channel dimension to obtain the nonlinear transformation branch feature F. dw1 Gated modulation branch characteristics F dw2 : ; in, This is used to divide the output features of depthwise separable convolution DWConv into two independent sub-tensors along the channel dimension. For the nonlinear transformation branch feature F dw1 Perform nonlinear activation of GELU and combine it with gated modulation branch features F dw2Element-wise multiplication is performed to achieve adaptive modulation of the features, and then projective convolution ProjConv is used to map the features back to the target channel dimension to obtain the projected enhanced features F. proj : ; S64. Introduce a spatial attention mechanism to enhance the projection feature F. proj Calculate the mean feature Mean(F) along the channel dimension. proj ), Maximum value feature Max(F) proj ), and through spatial convolution Conv s The spatial weight map is generated using the Sigmoid function to enhance the projected feature F. proj Weighted modulation is applied to further enhance the network's ability to focus on key spatial regions, strengthen texture-salient areas, and suppress redundant background responses, resulting in spatial attention features F. sattn : ; in, This indicates that the mean feature Mean(F) proj ) and the maximum value feature Max(F) proj (Synthesizing is performed along the channel dimension;) S65. In the output stage, a residual strategy is adopted to incorporate spatial attention features F. sattn Texture features F texture and color characteristics F color The process involves fusion to maintain the integrity of the original texture and color information, resulting in a fusion feature Y that effectively enhances local texture information while ensuring color consistency. .

[0012] (III) Beneficial Effects Compared with existing technologies, the image quality optimization method based on multi-branch collaboration provided by this invention has the following beneficial effects: 1) A mechanism for jointly processing texture restoration and color correction is proposed, which breaks through the technical bottleneck of "mutual constraint between color correction and detail preservation" in traditional methods. By co-modeling and jointly optimizing local texture information and global color information within a unified framework, it effectively avoids the problems of edge blurring and loss of details caused by emphasizing color correction alone in existing technologies, thereby improving color accuracy while fully maintaining the integrity of image structure and texture. 2) A dual-branch structure of local texture enhancement module and global color enhancement module was designed. The local texture enhancement module focuses on capturing high-frequency details and structural information of the image, while the global color enhancement module is used to model the relationship between the overall color distribution and hue. By specifically modeling local texture information and global color information, the ability to express complex image features is effectively improved, providing a richer and more complementary information foundation for subsequent feature fusion. 3) Further, an attention-guided feature fusion module is introduced to adaptively weight and integrate texture and color features. This fusion mechanism can dynamically adjust the contribution ratio of each branch information according to the importance of different channel features and spatial positions, thereby achieving a more reasonable image enhancement strategy in different regions. This effectively improves the overall performance of the final sRGB image in terms of detail preservation, color naturalness, and overall visual quality. Through local-global feature branches and attention-guided adaptive fusion, it effectively compensates for the information loss caused by ISP compression, while maintaining dark details and color reproduction. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0014] Figure 1 This is a schematic diagram of the network structure of the present invention; Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] The core of this invention lies in the introduction of local information enhancement, global information enhancement, and texture enhancement branches into the existing DNF framework. This achieves joint optimization of image brightness, detail texture, and color consistency, effectively improving the visual quality of images in low-light and noisy environments. The overall network structure (such as...) Figure 1As shown, the improved U-Net architecture is adopted, which achieves decoupled processing of different types of visual degradation through front-end light amplification, encoder feature extraction, multi-branch enhancement and gating fusion design.

[0017] Unlike existing methods that rely solely on single-path features or fixed post-processing strategies, this invention achieves targeted quality improvement through multi-branch collaborative enhancement and intelligent fusion mechanisms. In optional implementations, macroscopic visual feedback is introduced to form a system-level closed loop, resulting in stronger adaptability and generalization capabilities.

[0018] The following describes the specific process of the image quality optimization method based on multi-branch collaboration provided by this invention, using specific examples (e.g.) Figure 2 (as shown) and technical effects.

[0019] S1. Perform light amplification processing on the input low-light image, including: To address underexposure caused by low-light environments, the input low-light image I is enhanced using a linear transformation combined with adaptive brightness adjustment, resulting in the enhanced low-light image I. light : ;

[0020] ; Where ratio is the exposure magnification factor. , These are the exposure times for normal light images and low light images, respectively.

[0021] S2. Input the enhanced low-light image into the denoising decoding branch to effectively suppress noise and preserve image structural information, obtaining denoising features, including: S21, Enhance the low-light image I light Input encoder to extract initial features F enc : ; S22, Regarding the initial feature F enc Denoising and decoding are performed to obtain the denoised features F. denoised .

[0022] S3. Input the denoising features into the local information enhancement branch. While maintaining the stability of the original structure of the low-light image, enhance the local texture information in the dark areas, providing local enhancement features for subsequent feature fusion and image reconstruction, including: S31. Considering the severe loss of texture information in dark areas, an adaptive convolutional kernel prediction mechanism is introduced to specifically enhance the details in high-noise and low-brightness areas, resulting in the basic convolutional feature F. con : ; Among them, the basic convolutional feature F con By analyzing the denoising feature F denoised Perform two 3x3 convolutions With nonlinear activation This is obtained after performing local context modeling and feature reconstruction; S32. Introduce a channel attention mechanism to adaptively model the importance of different channels, and apply this to the denoising feature F. denoised Global average pooling (GAP) is performed to obtain channel-level statistics, and channel weights are generated using a multilayer perceptron (MLP) and a sigmoid function to optimize the basic convolutional features. Channel-by-channel recalibration is performed to highlight channels that are more critical for restoring texture information in dark areas, while suppressing redundant and noisy channels, resulting in the channel recalibration feature F. ca : ; in, The Hadamard product represents element-wise multiplication. S33. Introduce a spatial attention mechanism by recalibrating the channel features F. ca Average pooling (AvgPool) and max pooling (MaxPool) are performed separately to aggregate spatial dimensionality information, and then a 7x7 convolution is applied. The spatial weight map is generated using the Sigmoid function, and the channel features F are recalibrated. ca Adaptive enhancement is performed in spatial location to strengthen the effective texture structure in high-noise and low-brightness areas and improve the discernibility of texture information in dark areas, resulting in spatial recalibration features F. sa : ; S34. Spatial recalibration feature F sa With denoising feature F denoised Perform residual fusion and then use 1*1 convolution. Channel compression and feature integration are performed to obtain the local enhanced feature F. local : .

[0023] S4. Input the denoising features into the global information enhancement branch. Addressing the differences in overall brightness distribution and color consistency in low-light images, perform global color information modeling and compensation on the denoising features to obtain color features, including: S41. Denoising feature F is achieved by using multi-layer convolution and non-linear activation LeakyReLU. denoised Initial mapping is performed, and the denoised features F are processed through local receptive field convolution. denoisedSmoothing and recoding are performed to extract globally enhanced intermediate features F that are beneficial for modeling overall brightness and color distribution. Re : ; S42, Based on globally enhanced intermediate features F Re The generated channel weights m(F) Re The weighted values ​​are then processed, and the weighted results are subjected to global average pooling (GAP) in the spatial dimension to obtain statistical features reflecting the overall brightness and color distribution. Finally, a 1x1 convolution is applied. The Sigmoid function generates global context compensation information Context(F) Re ): ; S43. Transfer global context compensation information Context(F) Re Globally enhanced intermediate features F Re and denoising features F denoised By adding elements one by one, we obtain the color feature F, which integrates the overall brightness distribution and color consistency information. color : .

[0024] S5. Input the denoising features into the texture enhancement branch. By combining explicit texture guidance and implicit attention modeling, the feature enhancement mechanism targets and enhances high-frequency structural information to recover the detail edges and local texture information lost in the low-light image after denoising. This improves the clarity and visual detail of the final reconstructed image, resulting in texture features, including: S51, Regarding the denoising feature F denoised Calculate the Sobel gradient to further recover detailed edges and local texture information, and obtain the gradient magnitude G: ; in, , These are the Sobel gradient operators in the X and Y directions, respectively. It is a positive number that approaches 0, used to avoid numerical instability when the gradient is zero; Normalizing the gradient magnitude G eliminates the influence of gradient magnitude scale differences between different images or regions, resulting in texture gradient features that are more suitable as structural prior information for subsequent attention calculations while maintaining the relative texture distribution. : ; Where max(G) is the maximum value in the gradient magnitude G; S52, Texture gradient features With denoising feature F denoised The system concatenates data along the channel dimension to construct a joint feature representation containing local texture and semantic information. Pointwise convolution (PWConv) is then used for channel mapping and dimensionality compression. Finally, depthwise separable convolution (DWConv) is employed to extract local spatial relationships, generating a query vector Q, a key vector K, and a value vector V for the attention mechanism. ; in, This indicates that the texture gradient feature With denoising feature F denoised The components are stitched together along the channel dimension. S53. Calculate the attention weight matrix Attn along the channel dimension to measure the importance of different channels in the texture enhancement task, and obtain the texture-aware enhancement feature F by weighted summation. attn This allows the network to adaptively focus on the channels most critical for texture restoration, suppressing ineffective or noise-dominated effects. ; ; in, This is a temperature coefficient used to adjust the smoothness of attention distribution. The Softmax function is used to ensure that the attention weights have interpretable probability distribution characteristics; S54. Integrate gradient magnitude G with texture-aware enhancement feature F attn By fusing explicit texture guidance and implicit attention modeling, the system effectively enhances realistic edges and texture structures while avoiding over-sharpening, resulting in the fused enhanced feature F. enh : ; in, This indicates a convolution mapping of the gradient magnitude G. These are the weighting coefficients. ; S55. A feedforward network is introduced at the end of the texture enhancement branch, and then connected to the denoising feature F through residual connections. denoised The features are fused to further enhance their representational power and stabilize the training process, resulting in texture features F. texture : .

[0025] S6. Input texture features and color features into the fusion module. Through multi-level attention modeling and convolutional transformation, co-model and effectively fuse local texture information and global color information to obtain fused features that effectively enhance local texture information while ensuring color consistency, including: S61, Considering texture features F texture With color feature F color The importance of different channel dimensions varies significantly. First, channel attention modeling is performed on the two types of features, and then for the texture feature F... texture Global average pooling (GAP) is performed to obtain channel-level statistics, and then channel convolution (Conv) is applied. c The Sigmoid function generates channel weights for the texture feature F. texture Weighted modulation is performed to obtain the texture attention feature F' texture : ; For color feature F color Using texture feature F texture Using the same channel attention modeling method, we obtain the color attention feature F'. color : ; Through the above processing, the network is able to adaptively emphasize channels that are more critical to texture restoration or color consistency; S62, Apply texture attention feature F' texture With color attention feature F' color The data is concatenated along the channel dimension, and then channel recombination and dimensionality compression are performed using pointwise convolution PWConv to obtain the intermediate feature F. pw : ; S63. To further model local spatial relationships, a depthwise separable convolution (DWConv) is introduced to model intermediate features F. pw The transformation is performed, splitting the output features into two parts along the channel dimension to obtain the nonlinear transformation branch feature F. dw1 Gated modulation branch characteristics F dw2 : ; in, This is used to divide the output features of depthwise separable convolution DWConv into two independent sub-tensors along the channel dimension. For the nonlinear transformation branch feature F dw1 Perform nonlinear activation of GELU and combine it with gated modulation branch features F dw2Element-wise multiplication is performed to achieve adaptive modulation of the features, and then projective convolution ProjConv is used to map the features back to the target channel dimension to obtain the projected enhanced features F. proj : ; S64. Introduce a spatial attention mechanism to enhance the projection feature F. proj Calculate the mean feature Mean(F) along the channel dimension. proj ), Maximum value feature Max(F) proj ), and through spatial convolution Conv s The spatial weight map is generated using the Sigmoid function to enhance the projected feature F. proj Weighted modulation is applied to further enhance the network's ability to focus on key spatial regions, strengthen texture-salient areas, and suppress redundant background responses, resulting in spatial attention features F. sattn : ; in, This indicates that the mean feature Mean(F) proj ) and the maximum value feature Max(F) proj (Synthesizing is performed along the channel dimension;) S65. In the output stage, a residual strategy is adopted to incorporate spatial attention features F. sattn Texture features F texture and color characteristics F color The process involves fusion to maintain the integrity of the original texture and color information, resulting in a fusion feature Y that effectively enhances local texture information while ensuring color consistency. .

[0026] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image quality optimization method based on multi-branch collaboration, characterized in that: Includes the following steps: S1. Perform light magnification processing on the input low-light image; S2. Input the enhanced low-light image into the denoising decoding branch to effectively suppress noise and preserve image structure information, thereby obtaining denoising features; S3. Input the denoising features into the local information enhancement branch. While maintaining the stability of the original structure of the low-light image, enhance the local texture information in the dark area and provide local enhancement features for subsequent feature fusion and image reconstruction. S4. Input the denoising features into the global information enhancement branch. Based on the differences in the overall brightness distribution and color consistency of the low-light image, perform global color information modeling and compensation on the denoising features to obtain color features. S5. Input the denoising features into the texture enhancement branch. By combining the feature enhancement mechanism of explicit texture guidance and implicit attention modeling, the high-frequency structural information is specifically enhanced to restore the detail edges and local texture information lost in the low-light image after denoising, thereby improving the clarity and visual detail of the final reconstructed image and obtaining texture features. S6. Input texture features and color features into the fusion module. Through multi-level attention modeling and convolution transformation, the local texture information and global color information are collaboratively modeled and effectively fused to obtain fusion features that effectively enhance local texture information while ensuring color consistency.

2. The image quality optimization method based on multi-branch collaboration according to claim 1, characterized in that: S1 performs light amplification processing on the input low-light image, including: To address underexposure caused by low-light environments, the input low-light image I is enhanced using a linear transformation combined with adaptive brightness adjustment, resulting in the enhanced low-light image I. light : ； ； Where ratio is the exposure magnification factor. , These are the exposure times for normal light images and low light images, respectively.

3. The image quality optimization method based on multi-branch collaboration according to claim 2, characterized in that: In S2, the enhanced low-light image is input into the denoising decoding branch, effectively suppressing noise while preserving image structural information, resulting in denoising features, including: S21, Enhance the low-light image I light Input encoder to extract initial features F enc : ； S22, Regarding the initial feature F enc Denoising and decoding are performed to obtain the denoised features F. denoised .

4. The image quality optimization method based on multi-branch collaboration according to claim 3, characterized in that: In S3, the denoising features are input into the local information enhancement branch. While maintaining the stability of the original structure of the low-light image, the local texture information in the dark areas is enhanced, providing local enhancement features for subsequent feature fusion and image reconstruction, including: S31. Considering the severe loss of texture information in dark areas, an adaptive convolutional kernel prediction mechanism is introduced to specifically enhance the details in high-noise and low-brightness areas, resulting in the basic convolutional feature F. con : ； Among them, the basic convolutional feature F con By analyzing the denoising feature F denoised Perform two 3x3 convolutions With nonlinear activation This is obtained after performing local context modeling and feature reconstruction; S32. Introduce a channel attention mechanism to adaptively model the importance of different channels, and apply this to the denoising feature F. denoised Global average pooling (GAP) is performed to obtain channel-level statistics, and channel weights are generated using a multilayer perceptron (MLP) and a sigmoid function to optimize the basic convolutional features. Channel-by-channel recalibration is performed to highlight channels that are more critical for restoring texture information in dark areas, while suppressing redundant and noisy channels, resulting in the channel recalibration feature F. ca : ； in, The Hadamard product represents element-wise multiplication. S33. Introduce a spatial attention mechanism by recalibrating the channel features F. ca Average pooling (AvgPool) and max pooling (MaxPool) are performed separately to aggregate spatial dimensionality information, and then a 7x7 convolution is applied. The spatial weight map is generated using the Sigmoid function, and the channel features F are recalibrated. ca Adaptive enhancement is performed in spatial location to strengthen the effective texture structure in high-noise and low-brightness areas and improve the discernibility of texture information in dark areas, resulting in spatial recalibration features F. sa : ； S34. Spatial recalibration feature F sa With denoising feature F denoised Perform residual fusion and then use 1*1 convolution. Channel compression and feature integration are performed to obtain the local enhanced feature F. local : 。 5. The image quality optimization method based on multi-branch collaboration according to claim 4, characterized in that: In S4, the denoising features are input into the global information enhancement branch. Addressing the differences in overall brightness distribution and color consistency in low-light images, global color information modeling and compensation are performed on the denoising features to obtain color features, including: S41. Denoising feature F is achieved by using multi-layer convolution and non-linear activation LeakyReLU. denoised Initial mapping is performed, and the denoised features F are processed through local receptive field convolution. denoised Smoothing and recoding are performed to extract globally enhanced intermediate features F that are beneficial for modeling overall brightness and color distribution. Re : ； S42, Based on globally enhanced intermediate features F Re The generated channel weights m(F) Re The weighted values ​​are then processed, and the weighted results are subjected to global average pooling (GAP) in the spatial dimension to obtain statistical features reflecting the overall brightness and color distribution. Finally, a 1x1 convolution is applied. The Sigmoid function generates global context compensation information Context(F) Re ): ； S43. Transfer global context compensation information Context(F) Re Globally enhanced intermediate features F Re and denoising features F denoised By adding elements one by one, we obtain the color feature F, which integrates the overall brightness distribution and color consistency information. color : 。 6. The image quality optimization method based on multi-branch collaboration according to claim 5, characterized in that: In S5, denoising features are input into the texture enhancement branch. By combining explicit texture guidance and implicit attention modeling, a feature enhancement mechanism is used to specifically enhance high-frequency structural information, thereby restoring the detailed edges and local texture information lost in the low-light image after denoising. This improves the clarity and visual detail of the final reconstructed image, resulting in texture features, including: S51, Regarding the denoising feature F denoised Calculate the Sobel gradient to further recover detailed edges and local texture information, and obtain the gradient magnitude G: ； in, , These are the Sobel gradient operators in the X and Y directions, respectively. It is a positive number that approaches 0, used to avoid numerical instability when the gradient is zero; Normalizing the gradient magnitude G eliminates the influence of gradient magnitude scale differences between different images or regions, resulting in texture gradient features that are more suitable as structural prior information for subsequent attention calculations while maintaining the relative texture distribution. : ； Where max(G) is the maximum value in the gradient magnitude G; S52, Texture gradient features With denoising feature F denoised The system concatenates data along the channel dimension to construct a joint feature representation containing local texture and semantic information. Pointwise convolution (PWConv) is then used for channel mapping and dimensionality compression. Finally, depthwise separable convolution (DWConv) is employed to extract local spatial relationships, generating a query vector Q, a key vector K, and a value vector V for the attention mechanism. ； in, This indicates that the texture gradient feature With denoising feature F denoised The components are stitched together along the channel dimension. S53. Calculate the attention weight matrix Attn along the channel dimension to measure the importance of different channels in the texture enhancement task, and obtain the texture-aware enhancement feature F by weighted summation. attn This allows the network to adaptively focus on the channels most critical for texture restoration, suppressing ineffective or noise-dominated effects. ； ； in, This is a temperature coefficient used to adjust the smoothness of attention distribution. The Softmax function is used to ensure that the attention weights have interpretable probability distribution characteristics; S54. Integrate gradient magnitude G with texture-aware enhancement feature F attn By fusing explicit texture guidance and implicit attention modeling, the system effectively enhances realistic edges and texture structures while avoiding over-sharpening, resulting in the fused enhanced feature F. enh : ； in, This indicates a convolution mapping of the gradient magnitude G. These are the weighting coefficients. ; S55. A feedforward network is introduced at the end of the texture enhancement branch, and then connected to the denoising feature F through residual connections. denoised The features are fused to further enhance their representational power and stabilize the training process, resulting in texture features F. texture : 。 7. The image quality optimization method based on multi-branch collaboration according to claim 6, characterized in that: In S6, texture and color features are input into the fusion module. Through multi-level attention modeling and convolutional transformation, local texture information and global color information are collaboratively modeled and effectively fused to obtain fused features that effectively enhance local texture information while ensuring color consistency. These features include: S61, Considering texture features F texture With color feature F color The importance of different channel dimensions varies significantly. First, channel attention modeling is performed on the two types of features, and then for the texture feature F... texture Global average pooling (GAP) is performed to obtain channel-level statistics, and then channel convolution (Conv) is applied. c The Sigmoid function generates channel weights for the texture feature F. texture Weighted modulation is performed to obtain the texture attention feature F' texture : ； For color feature F color Using texture feature F texture Using the same channel attention modeling method, we obtain the color attention feature F'. color : ； Through the above processing, the network is able to adaptively emphasize channels that are more critical to texture restoration or color consistency; S62, Apply texture attention feature F' texture With color attention feature F' color The data is concatenated along the channel dimension, and then channel recombination and dimensionality compression are performed using pointwise convolution PWConv to obtain the intermediate feature F. pw : ； S63. To further model local spatial relationships, a depthwise separable convolution (DWConv) is introduced to model intermediate features F. pw The transformation is performed, splitting the output features into two parts along the channel dimension to obtain the nonlinear transformation branch feature F. dw1 Gated modulation branch characteristics F dw2 : ； in, This is used to divide the output features of depthwise separable convolution DWConv into two independent sub-tensors along the channel dimension. For the nonlinear transformation branch feature F dw1 Perform nonlinear activation of GELU and combine it with gated modulation branch features F dw2 Element-wise multiplication is performed to achieve adaptive modulation of the features, and then projective convolution ProjConv is used to map the features back to the target channel dimension to obtain the projected enhanced features F. proj : ； S64. Introduce a spatial attention mechanism to enhance the projection feature F. proj Calculate the mean feature Mean(F) along the channel dimension. proj ), Maximum value feature Max(F) proj ), and through spatial convolution Conv s The spatial weight map is generated using the Sigmoid function to enhance the projected feature F. proj Weighted modulation is applied to further enhance the network's ability to focus on key spatial regions, strengthen texture-salient areas, and suppress redundant background responses, resulting in spatial attention features F. sattn : ； in, This indicates that the mean feature Mean(F) proj ) and the maximum value feature Max(F) proj (Synthesizing is performed along the channel dimension;) S65. In the output stage, a residual strategy is adopted to incorporate spatial attention features F. sattn Texture features F texture and color characteristics F color The process involves fusion to maintain the integrity of the original texture and color information, resulting in a fusion feature Y that effectively enhances local texture information while ensuring color consistency. 。

Images