High-definition image enhancement method based on multi-scale frequency characteristic optimization and dynamic fusion

Abstract

The invention relates to a high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion, which comprises the following steps: firstly, carrying out high-low frequency decoupling processing on an image, respectively enhancing an obtained low-frequency component and an obtained high-frequency component, and then carrying out feature integration on the low-frequency component L3 before and after enhancement to generate an information guide feature; information guide features are subjected to up-sampling by two times, are added with features which are processed by a high-frequency detail preliminary enhancement module and are extracted through 3 * 3 convolution, and are input into a first-level feature fusion module; and finally, the finally enhanced high-frequency component of the layer is obtained through 3 * 3 convolution and is reconstructed into a complete enhanced image to be output according to a similar method. Through the synergistic effect of the enhancement module and the hierarchical feature fusion module, the brightness, color and high-frequency details of the high-definition image can be effectively recovered and enhanced, and meanwhile, the problems of artifacts and noise amplification are avoided.

Description

Technical Field

[0001] The present invention relates to a high-definition image enhancement method, and in particular to a high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion. Background Art

[0002] With the continuous upgrades of cameras and smartphones, people's pursuit of image quality is increasing, and photography has become an important way to record and share daily life. However, in complex real-world environments, image quality is often affected by a variety of factors. In particular, under suboptimal lighting conditions, the resulting images may suffer from insufficient brightness, loss of detail, and low contrast. These issues not only reduce the aesthetic quality of the images but also seriously hinder the performance of advanced vision tasks such as object detection, recognition, tracking, image segmentation, and scene understanding.

[0003] To improve low-quality images, traditional methods often rely on professional retouchers to manually adjust parameters such as brightness, contrast, and color to enhance visual perception and optimize image processing tasks. However, this approach suffers from low efficiency and inconsistent quality, making it difficult to meet the needs of large-scale applications. With the development of computer vision technology, automated image enhancement methods have gradually become a research hotspot, aiming to reduce reliance on manual adjustments and improve processing stability and consistency.

[0004] Most existing models focus on enhancing brightness and contrast, such as histogram equalization and Retinex theory, but typically fail to consider processing multiple image features, such as global image features like brightness and color, as well as high-frequency details like edges. Furthermore, when processing high-definition images, most models downsample the image before inputting it into the model for learning to reduce the burden on computational resources. However, this approach can result in the loss of rich details and information in high-definition images, making the enhancement difficult to meet the human eye's perception requirements. Most image enhancement methods typically rely on complex network architectures to simultaneously address low- and high-frequency issues, essentially processing the image at a single scale. This approach ignores the coupling between high- and low-frequency information, and due to frequency optimization issues, it often leads to mutual interference between low- and high-frequency information, compromising the final enhancement effect. For example, low-frequency adjustments can inadvertently amplify noise, while high-frequency restoration can affect light intensity recovery.

[0005] In summary, existing high-definition image enhancement technologies have the following main problems that need to be solved:

[0006] 1. Improper frequency information coupling: Existing technologies often use a single scale to process images, processing both low-frequency and high-frequency information simultaneously, ignoring the coupling relationship between high- and low-frequency information. As a result, low-frequency adjustments may amplify noise, while high-frequency recovery may affect light intensity.

[0007] 2. Insufficient multi-feature fusion: Most models focus on enhancing brightness and contrast, and fail to fully consider multiple image features, such as the comprehensive optimization of global features (brightness, color) and local features (high-frequency details such as edges).

[0008] 3. Loss of high-definition detail information: To reduce the burden on computing resources, existing models often downsample high-definition images before inputting them into the network. This results in the loss of rich detail information in the high-definition images, and the resulting enhancement effect is ultimately unable to meet the perception needs of the human eye. Summary of the Invention

[0009] In view of the above problems existing in the prior art, the technical problems to be solved by the present invention are: improper frequency information coupling processing, insufficient multi-feature fusion and loss of high-definition detail information in paired high-definition image enhancement technology.

[0010] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0011] A high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion includes the following steps:

[0012] S1: Image high and low frequency decoupling processing, the input image Decompose into low-frequency components and high-frequency components Among them, h and w represent the height and width of the original image respectively, and H0, H1, and H2 correspond to the high-frequency detail information extracted at different resolutions respectively.

[0013] S2: Enhancement module includes a global enhancement module for enhancing L3 and a module for enhancing high frequency components. A high-frequency detail preliminary enhancement module for detail enhancement;

[0014] For the low-frequency component L3, the 3x3 convolution layer in the global enhancement module increases the number of channels of L3 from 3 to 48, and then uses 5 attention state space modules to stack and extract high-level semantics. Then a 3x3 convolution is used to integrate the output deep features, and finally a 3x3 convolution is used to project the features back to the RGB three channels to output the enhanced low-frequency feature map.

[0015] The high-frequency detail preliminary enhancement module includes a feature enhancement branch, a feature enhancement branch, and a feature fusion mechanism;

[0016] Gradient detection branch: Use Sobel operator to extract The gradient information in the horizontal and vertical directions is combined and all-round edge features are obtained through additive fusion, and then channel integration is performed through 1×1 convolution.

[0017] Feature enhancement branch: consists of 3 residual blocks cascaded to extract The deep features are then transformed through 3x3 convolution, and the important areas are selectively strengthened through the spatial attention mechanism module. The feature representation is further optimized through two 1×1 convolution layers, and the final feature fusion and channel adjustment are performed using three series residual blocks and a 3x3 convolution.

[0018] Feature fusion mechanism: The output of the gradient detection branch and the output of the feature enhancement branch are fused in the channel dimension, and the features are projected back to the RGB three channels through 3×3 convolution to output the enhanced high-frequency feature map

[0019] S3: Connect L3 and It directly splices in the channel dimension, and passes through a 3x3 convolution block to increase the number of channels from 3 to 32. Then, three residual blocks are used to extract complex features, and then the spatial attention module is used to focus on important areas. Finally, a 3x3 convolution is used to integrate the features to generate information-guided features.

[0020] S4: The information-guided features are upsampled by 2 times and then The features extracted by 3x3 convolution are added and input to the first level feature fusion module K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. The output of K1 is upsampled by 2 times and then The features extracted by 3x3 convolution are added and input to the second level feature fusion module K2 adjacent to K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. Then upsample the output of K2 by 2 times and add Add them together and input them into the third level feature fusion module K3, and then go through a 3x3 convolution to get the final enhanced high-frequency component of this layer.

[0021] S5: and The Laplacian pyramid is used to reconstruct the complete enhanced image output I.

[0022] S6: Calculate I and the corresponding high-quality reference image I with uniform illumination, coordinated brightness and color high The loss between I and I high Perform pixel-by-pixel comparison, calculate the L1 loss value, and reversely update the parameters in S1-S5 based on the loss value. When the loss value no longer decreases, the training is completed and the current parameters in S1-S5 are retained.

[0023] S7: For a new image, the new image is processed by the current S1-S5, and the output is the corresponding enhanced image.

[0024] Furthermore, in S1, the Laplace pyramid technique is used to perform high- and low-frequency decoupling processing on the input image to obtain low-frequency components and high-frequency components.

[0025] Furthermore, the process of obtaining the corresponding features of the high-frequency components of the final enhancement in this layer by the mid-level feature fusion module in S4 is as follows:

[0026] Average pooling and maximum pooling are used to extract global statistical features from the input of the hierarchical feature fusion module. The two extracted global statistical features are nonlinearly transformed through the MLP network respectively, and then the results of the two nonlinear transformations are channel-wise added to obtain the global features of this layer.

[0027] The input of the hierarchical feature fusion module uses 1×1 convolution + ReLU to extract local features, and uses multi-scale depth-wise separable convolution 3×3DWConv, 5×5DWConv and 7×7DWConv to extract texture features under different receptive fields respectively, and then the three texture features are added together to obtain the local features of this layer.

[0028] The global features and local features of this layer are input into the adaptive feature fusion module for preliminary processing add, that is, the multi-scale features are aggregated through concat, the weights of the local features of this layer and the weights of the global features of this layer during aggregation are adjusted by the Sigmoid function, and the aggregated features are input into three series-connected residual modules for final feature fusion, so as to obtain the features corresponding to the final enhanced high-frequency components of this layer.

[0029] Compared with the prior art, the present invention has at least the following advantages:

[0030] 1. Through the synergistic effect of the high-definition detail preliminary enhancement module and the hierarchical feature fusion module, the present invention can effectively restore and enhance the problem of high-frequency detail loss in images caused by conditions such as low illumination. In particular, it shows excellent performance in preserving and strengthening fine structures such as edges and textures, making the enhanced image have higher clarity and visual expressiveness, while avoiding problems such as artifacts and noise amplification during the enhancement process.

[0031] This method decouples and processes the high- and low-frequency information in an image separately, combining it with multi-scale feature fusion and edge-preserving strategies. This effectively addresses the problems of existing paired HD image enhancement techniques, such as improper frequency information coupling, insufficient multi-feature fusion, and loss of HD detail information. This method achieves more efficient and high-quality HD image enhancement. This method demonstrates superior enhancement performance under a variety of complex lighting conditions, providing more reliable image input for advanced visual tasks.

[0032] 2. Enhanced effect quality improvement: Through high and low frequency separation processing, the present invention can effectively restore and enhance high-frequency detail information while ensuring global brightness and color adjustment, so that the enhanced image has higher clarity and visual expression.

[0033] 3. Improved processing efficiency: The adoption of a lightweight network structure and an efficient feature extraction module reduces computational complexity, improves processing speed while ensuring the enhancement effect, and is more suitable for practical application scenarios.

[0034] 4. Adaptability enhancement: Through the adaptive parameter adjustment mechanism, the present invention can automatically adjust the enhancement parameters according to the characteristics of the input image, so that the enhancement effect is more in line with the needs of different scenarios, thereby improving the adaptability and robustness of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This figure shows some experimental results on the Fivek dataset.

[0036] Figure 2 A visual comparison of the baseline method on the fivek dataset.

[0037] Figure 3 The following figure shows some experimental results on the private RCA dataset.

[0038] Figure 4 It is a simplified flow chart of the method of the present invention.

[0039] Figure 5 This is a simplified structural diagram of the high-frequency detail preliminary enhancement module.

[0040] Figure 6 The structure diagram of the hierarchical feature fusion module is shown in Figure 2. Detail enhancement implementation method DETAILED DESCRIPTION

[0041] The present invention is described in further detail below.

[0042] The overall processing process of the method of the present invention is: for the input image First, the Laplace pyramid is used to decompose the high and low frequency components to achieve effective decoupling of frequency information and obtain a set of high frequency components with gradual scale variation. And a low-frequency component L3 with the smallest scale. The low-frequency component is enhanced using the global enhancement module to obtain First, the three high-frequency components are input into the high-frequency detail preliminary enhancement module for preliminary detail enhancement. The low-frequency component L3 before enhancement and the enhanced Input to the information guidance module to obtain information guidance features, and then upsample the information guidance features by 2 times and combine them with The features extracted by 3x3 convolution are added and input into the hierarchical feature fusion module, and then another 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. The features output by the previous level feature fusion module are upsampled by 2 times and then The features extracted by 3x3 convolution are added and input into the hierarchical feature fusion module, and then another 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. Then the features output by the previous level feature fusion module are upsampled by 2 times and combined with Add them together and input them into the hierarchical feature fusion module, and then go through a 3x3 convolution to obtain the final enhanced high-frequency component of this layer. The high-frequency components of each layer are enhanced With the enhanced The Laplacian pyramid is used to reconstruct the complete enhanced image.

[0043] This invention effectively solves the problems existing in the existing paired high-definition image enhancement technology by innovatively decoupling the high- and low-frequency information of the image:

[0044] 1. Addressing frequency information coupling: By decomposing the input image into low-frequency and high-frequency components, dedicated processing paths are designed to enhance each separately, avoiding interference between high- and low-frequency information. The low-frequency processing path focuses on adjusting global brightness and color, while the high-frequency processing path focuses on enhancing details such as edges and textures.

[0045] 2. Addressing the issue of insufficient multi-feature fusion: We designed a high-definition detail preliminary enhancement module and a hierarchical feature fusion module, comprehensively considering multiple image features. The high-definition detail preliminary enhancement module extracts edge gradient features using the Sobel operator, combining a spatial attention mechanism with multiple residual blocks to effectively enhance high-frequency information. The hierarchical feature fusion module extracts and fuses features from different dimensions through operations such as average pooling, max pooling, and multi-scale depthwise separable convolution, achieving comprehensive optimization of global and local features.

[0046] 3. Addressing the issue of loss of high-definition detail information: A multi-scale feature pyramid structure is employed to extract and fuse image features at different resolution levels. Combined with a lightweight network design, this allows for direct processing of high-resolution images, avoiding information loss due to downsampling and preserving the rich details of the original high-definition image.

[0047] Example 1: A high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion, comprising the following steps:

[0048] S1: Image high and low frequency decoupling processing, the input image Decompose into low-frequency components and high-frequency components Here, h and w represent the height and width of the original image, respectively. H0, H1, and H2 correspond to high-frequency detail information extracted at different resolutions, with sizes of h×w, h / 2×w / 2, and h / 4×w / 4, respectively. They are primarily used to capture fine-grained structural features such as edges and textures. The low-frequency component L3 represents the image's global brightness, color, and structure, with the lowest spatial resolution but the richest semantics. These components are enhanced separately through different processing paths to avoid interference between high- and low-frequency information. The low-frequency component primarily preserves global information such as overall brightness and color, while the high-frequency component contains detailed information such as edges and textures.

[0049] Decomposition: For the input image First, the Laplace pyramid is used to decompose the high and low frequency components to achieve effective decoupling of frequency information and obtain a set of high frequency components with gradual scale variation. And a low-frequency component L3 with the smallest scale. The low-frequency component mainly stores global information such as the overall brightness and color of the image, while the high-frequency component contains detailed information such as edges and textures.

[0050] S2: Enhancement module includes a global enhancement module for enhancing L3 and a module for enhancing high frequency components. A preliminary high-frequency detail enhancement module for detail enhancement.

[0051] (1) The global enhancement module captures the overall brightness distribution and color information of the image through the global feature extraction module, and effectively corrects problems such as global illumination unevenness and color distortion. For the global enhancement module, specifically, for the input of low-frequency components A 3x3 convolution layer is used to increase the number of channels from 3 to 48 to meet the module's input requirements for channel dimensions. Five attention state space modules are then stacked to extract high-level semantic features. A 3x3 convolution is then used to integrate the output deep features. Finally, a 3x3 convolution is used to project the features back to the RGB channels and output the enhanced low-frequency components.

[0052] (2) The high-frequency detail preliminary enhancement module includes a feature enhancement branch, a feature enhancement branch, and a feature fusion mechanism;

[0053] The high-frequency components of the image are processed using the high-frequency detail preliminary enhancement module. This module uses the local feature extraction module to focus on strengthening the image's edge, texture and other detail information, and restores the high-frequency information loss caused by low illumination and other conditions. The module enhances the high-frequency details through the following structure:

[0054] Gradient detection branch: Sobel operator (Sobel x and Sobel y) is used to extract The gradient information in the horizontal and vertical directions is combined and all-round edge features are obtained through additive fusion, and then channel integration is performed through 1×1 convolution.

[0055] Feature enhancement branch: consists of 3 residual blocks cascaded to extract It uses deep features and selectively strengthens important areas through a spatial attention mechanism module, then further optimizes the feature representation through two 1×1 convolutional layers, and then uses three series residual blocks for final feature fusion and a 3x3 convolution for channel adjustment.

[0056] Feature fusion mechanism: The output of the gradient detection branch and the output of the feature enhancement branch are fused in the channel dimension, and the features are projected back to the RGB three channels through 3×3 convolution to output the enhanced high-frequency feature map

[0057] The attentional state-space module consists of two components. One is the Attentive State-space Equation (ASE), which leverages cue learning within the original Mamba state-space equation to query semantically similar pixels outside the scan sequence. Specifically, a learnable cue is designed to represent a set of pixels that are similar across the entire image. Representative cues are then injected into the output matrix C of the state-space equation via residual addition to derive the ASE. As a core component, ASE not only mitigates the causal nature of Mamba, improving performance, but also supports single-pass scanning, significantly improving efficiency. The other component is the Semantic Guided Neighboring (SGN), which encourages strong interactions between distant but similar pixels. Specifically, each pixel is first assigned a corresponding semantic label. Based on these labels, low-frequency components are then reconstructed to generate a semantically neighboring one-dimensional sequence where semantically similar pixels are also spatially close. By mitigating the long-range attenuation inherent in Mamba, SGN promotes effective interactions between distant but similar pixels in the original image.

[0058] A global enhancement module based on the attention state space model is designed to enhance the low-frequency components of the image. The global enhancement module captures the overall brightness distribution and color information of the image through the global feature extraction module, and effectively corrects problems such as global illumination unevenness and color distortion. For the global enhancement module, specifically, for the low-frequency components of the input image A 3x3 convolution layer is used to increase the number of channels from 3 to 48 to meet the module's input requirements for channel dimensions. Five attention state space modules are then stacked to extract high-level semantics, and a 3x3 convolution is used to integrate the output deep features. Finally, a 3x3 convolution is used to project the features back to the RGB channels and output the enhanced low-frequency feature map.

[0059] S3: Information guidance module: L3 and It directly splices in the channel dimension, and passes through a 3x3 convolution block to increase the number of channels from 3 to 32. Then, three residual blocks are used to extract complex features, and then the spatial attention module is used to focus on important areas. Finally, a 3x3 convolution is used to integrate the features to generate information-guided features.

[0060] S4: Hierarchical feature fusion module: A hierarchical feature fusion module is designed and used to implement bottom-up feature fusion to maintain the pixel correlation between the enhanced high-frequency component features and low-frequency component features, prevent artifacts, and reduce detail loss.

[0061] The information-guided features are upsampled by 2 times and then The features extracted by 3x3 convolution are added and input to the first level feature fusion module K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. The output of K1 is upsampled by 2 times and then The features extracted by 3x3 convolution are added and input to the second level feature fusion module K2 adjacent to K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. Then upsample the output of K2 by 2 times and add Add them together and input them into the third level feature fusion module K3, and then go through a 3x3 convolution to get the final enhanced high-frequency component of this layer. Through multi-scale feature extraction and adaptive fusion strategy, multi-level enhancement of high-frequency information is achieved: Taking the acquisition process of as an example, the process of obtaining the corresponding features of the high-frequency components of each layer by the hierarchical feature fusion module is described in detail:

[0062] Average pooling and maximum pooling are used to extract global statistical features from the input of the first-level feature fusion module. The two extracted global statistical features are nonlinearly transformed through the MLP network respectively, and then the results of the two nonlinear transformations are channel-wise added to obtain the first-layer global features.

[0063] The input of the first-level feature fusion module uses 1×1 convolution + ReLU to extract local features, and uses multi-scale depth-separable convolution 3×3DWConv, 5×5DWConv and 7×7DWConv to extract texture features under different receptive fields respectively, and then the three texture features are added together to obtain the first-level local features.

[0064] The first-layer global features and the first-layer local features are input into the adaptive feature fusion module for preliminary processing. That is, the multi-scale features are aggregated by concat, and the weights of the local features of this layer and the weights of the global features of this layer are adjusted by the Sigmoid function. After the aggregated features are input into three series-connected residual modules, the high-frequency components of the first layer are finally enhanced. Corresponding features.

[0065] S6: Calculate I and the corresponding high-quality reference image I with uniform illumination, coordinated brightness and color high The loss between I and I high Perform a pixel-by-pixel comparison and calculate the L1 loss (i.e., the mean pixel-level absolute error). This loss function effectively measures the overall visual difference between the generated image and the target image. By minimizing this error, backpropagating gradients and updating model parameters, the network is guided to continuously optimize its enhancement strategy during training, improving the brightness restoration, color fidelity, and detail preservation of the final output image. The parameters in S1-S5 are updated in reverse based on the loss value. When the loss value no longer decreases, training is complete and the current parameters in S1-S5 are retained.

[0066] S7: For a new image, the new image is processed by the current S1-S5 (i.e. S1, S2, S3, S4 and S5 when training is completed), and the output is the corresponding enhanced image.

[0067] S5: and The Laplacian pyramid is reconstructed into a complete enhanced image output.

[0068] In the field of image processing and computer vision, "important areas" usually refer to areas with high information content or discriminant value for the overall representation of the image or a specific task, mainly including:

[0069] High-contrast areas: Parts of an image where light and dark changes dramatically. These areas usually contain edge and contour information and are very important for object recognition and scene understanding.

[0070] Texture-rich regions: regions with complex texture patterns, which usually contain important information about the surface properties of objects.

[0071] Semantic key regions: Parts that are crucial to understanding the content of an image, such as the eyes, nose, and mouth in a face, or the main objects in a scene.

[0072] Abnormal areas: Areas that are significantly different from the surrounding environment. These may be targets or abnormal phenomena that require special attention.

[0073] Detail-rich regions: Regions containing fine structures are particularly important in image enhancement tasks because preserving and enhancing these regions can significantly improve visual quality.

[0074] Task-related areas: vary depending on the specific task. For example, in face beautification, skin texture and tone are important areas; in landscape photos, the sky and horizon may be important areas.

[0075] Experimental analysis:

[0076] 1. Experimental Setup

[0077] Dataset: The experiment uses multiple standard datasets including the public MIT-Adobe FiveK and the private RCA dataset, which contain paired images under different lighting conditions.

[0078] Evaluation indicators: Quantitative and qualitative indicators such as PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), LPIPS (Learned Perceptual Image Patch Similarity) and user study are used for evaluation.

[0079] Comparison method: KinD, DSLR, Uformer, Restormer, 4DLUT and other current mainstream image enhancement methods are selected for comparison.

[0080] 2. Experimental Results

[0081] (1) Comparison of objective indicators: On all test datasets, the method of the present invention outperforms the comparison method in terms of PSNR and SSIM indicators. For example, on the public high-definition dataset MIT-Adobe FiveK, the PSNR of the present invention reaches 24.9175dB and the SSIM reaches 0.9231, which are 0.404dB and 0.002 higher than the suboptimal method, respectively. On the dataset, the PSNR and SSIM of the present invention reach 27.31dB and 0.9252, respectively, as shown in Table 1.

[0082] Table 1 Average PSNR / SSIM / LPIPS indicators of test set images

[0083]

[0084] (2) Visual Quality Comparison: In terms of visual quality, the method of the present invention can better restore image details while maintaining natural colors and reasonable contrast. In particular, when processing low-light images of complex scenes, the present invention can effectively avoid problems such as over-enhancement and noise amplification, maintaining the naturalness and visual comfort of the image.

[0085] (3) Computational efficiency: Although the present invention adopts a decoupled processing strategy, it maintains high processing efficiency by optimizing the network structure and computational process. On standard GPU hardware, the present invention can process 1080p resolution images at 20 frames per second, meeting the requirements of real-time applications.

[0086] pass Figure 1 As can be seen, the model generates images of excellent quality, closely matching the target image in terms of overall color consistency and detail restoration. Furthermore, both foreground subjects and background elements exhibit high clarity and layering, with rich and natural details.

[0087] pass Figure 2As can be seen, the proposed model demonstrates high quality in terms of color, content restoration, and detail preservation. The first column shows the input image, which is generally darker, making many details difficult to discern. The second column shows the enhancement results using the KinD network. Its main problem lies in excessive brightness enhancement, resulting in color distortion. For example, the color of the collar in the second image is significantly shifted, and the hair detail within the rectangular box is significantly lost, indicating that KinD's processing of highlight areas is inaccurate. The third column shows the enhancement results using the DLSR network. This method exhibits severe color shift in some images, accompanied by significant noise pollution, which affects the enhancement quality. The fourth column shows the enhancement results using the Uformer network. Compared to earlier image enhancement networks, the Uformer network significantly improves overall enhancement quality, likely due to the Transformer architecture's advantage in capturing long-range dependencies. However, observing the details in the second image, the colors are significantly darker than those in the target image, indicating that its enhancement still has some shortcomings. The fifth column shows the enhancement results using the Restormer network, which achieves high overall quality, with relatively accurate colors and a pleasant aesthetic after enhancement. However, compared with the target image, Restormer still has certain deviations in brightness adjustment. This may be because as a unified network for image restoration, its learning ability for brightness adjustment is relatively limited. The sixth column is the enhancement result of the 4DLUT network. This method uses a four-dimensional lookup table (4D LUT) for color enhancement, which can effectively learn the nonlinear color mapping relationship between the input image and the target image. The overall quality of the enhanced image is relatively good, close to the target image, the color transition is natural, and the detail texture is relatively fine. However, there is still a certain color deviation in the local details of the second image. This may be because the 4D LUT is difficult to accurately adjust the local details when processing the global color mapping, resulting in slight deviations in the color consistency of some areas. The seventh column is the enhancement result of the model proposed by the present invention. High-quality enhancement effects are maintained from the overall to the local details. This is mainly due to the fact that the method proposed by the present invention adopts a high- and low-frequency decomposition strategy, and designs special enhancement modules for different frequency domains, avoiding the problem of information mixing in the single-scale enhancement process, thereby achieving a more reasonable enhancement effect. Specifically, this method not only ensures that the overall brightness and color of the image are consistent with the target Figure 1 At the same time, it also effectively enhances the details to make the image more natural and realistic.

[0088] pass Figure 3 It can be seen that the processing results of the method proposed in this patent in portrait photography have achieved good results in detail preservation, color correction, exposure balance, etc.

[0089] In summary, the multi-feature fusion and pairing high-definition image enhancement method based on decoupling of high- and low-frequency information of the image proposed in the present invention effectively solves the problems of improper frequency information coupling processing, insufficient multi-feature fusion, and loss of high-definition detail information in the existing technology, and achieves a more efficient and high-quality high-definition image enhancement effect, providing more reliable image input for advanced visual tasks, and has important theoretical value and application prospects.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion, characterized by: The steps include: S1: Image high and low frequency decoupling processing, the input image Decompose into low-frequency components and high-frequency components Where h and w represent the height and width of the original image, respectively. G0, G1, and G2 correspond to the high-frequency detail information extracted at different resolutions. S2: Enhancement module includes a global enhancement module for enhancing L3 and a module for enhancing high frequency components. A high-frequency detail preliminary enhancement module for detail enhancement; For the low-frequency component L3, the 3x3 convolution layer in the global enhancement module increases the number of L3 channels from 3 to 48, and then uses 5 attention state space modules to stack and extract high-level semantics, and then uses a 3×3 convolution to integrate and output deep features. Finally, a 3×3 convolution is used to project the deep features back to the RGB three channels to obtain the enhanced low-frequency feature map. The high-frequency detail preliminary enhancement module includes a gradient detection branch, a feature enhancement branch, and a feature fusion mechanism; Gradient detection branch: Use Sobel operator to extract Gradient information in the horizontal and vertical directions is added and fused to obtain all-round edge features, and then channel integration is performed through 1×1 convolution; Feature enhancement branch: consists of 3 residual blocks cascaded to extract Deep features, then channel transformation through 3x3 convolution, and then selectively enhance important areas through the spatial attention mechanism, followed by further optimization of feature representation through two 1x1 convolution layers, and then final feature fusion and channel adjustment using three series residual blocks and a 3x3 convolution; Feature fusion mechanism: The output of the gradient detection branch and the output of the feature enhancement branch are fused in the channel dimension, and the features are projected back to the RGB three channels through 3×3 convolution to output the enhanced high-frequency feature map S3: Connect L3 and Directly splicing in the channel dimension, passing through a 3x3 convolution block to increase its channel number from 3 to 32, then using three series residual blocks to extract complex features, and then passing through the spatial attention module to focus on important areas, and finally using 3x3 convolution to integrate features and generate information-guided features; S4: The information-guided features are upsampled by 2 times and then The features extracted by 3x3 convolution are added and input to the first level feature fusion module K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. The output of K1, that is, the feature corresponding to the high-frequency component of this layer, is upsampled by 2 times and then combined with The features extracted by 3x3 convolution are added and input to the second level feature fusion module K2 adjacent to K1, and then a 3x3 convolution is performed to obtain the final enhanced high-frequency component of this layer. Then the output of K2, that is, the feature corresponding to the high-frequency component of this layer, is upsampled by 2 times and combined with Add them together and input them into the third level feature fusion module K3, and then go through a 3x3 convolution to get the final enhanced high-frequency component of this layer. S5: and After the Laplacian pyramid is reconstructed into a complete enhanced image output I; S6: Calculate I and the corresponding high-quality reference image I with uniform illumination, coordinated brightness and color high The loss between I and I high Perform pixel-by-pixel comparison and calculate I and I high The L1 loss value between them is updated in reverse according to the loss value. When the loss value no longer decreases, the training is completed and the current parameters in S1-S5 are retained. S7: For a new image, the new image is processed by the current S1-S5, and the output is the corresponding enhanced image.

2. The high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion according to claim 1, characterized in that: In S1, the Laplace pyramid technique is used to perform high- and low-frequency decoupling processing on the input image to obtain low-frequency components and high-frequency components.

3. The high-definition image enhancement method based on multi-scale frequency feature optimization and dynamic fusion according to claim 1, characterized in that: The process of obtaining the features corresponding to the high-frequency components of this layer in S4 is as follows: Average pooling and maximum pooling are used to extract global statistical features from the input of the hierarchical feature fusion module. The two extracted global statistical features are nonlinearly transformed through the MLP network respectively, and then the results of the two nonlinear transformations are channel-wise added to obtain the global features of this layer. The input of the hierarchical feature fusion module uses 1×1 convolution + ReLU to extract local features, and uses multi-scale depth-separable convolution 3×3DWConv, 5×5DWConv and 7×7DWConv to extract texture features under different receptive fields respectively. Then the three texture features are added together to obtain the local features of this layer; The global features and local features of this layer are input into the adaptive feature fusion module for preliminary processing add, that is, the multi-scale features are aggregated through concat, the weights of the local features of this layer and the weights of the global features of this layer during aggregation are adjusted by the Sigmoid function, and the aggregated features are input into three series-connected residual modules for final feature fusion, so as to obtain the features corresponding to the high-frequency components of this layer.

Images