Low-illumination image enhancement method based on bidirectional cross-frequency-domain guided wavelet diffusion

Abstract

The invention discloses a low-illumination image enhancement method based on bidirectional cross-frequency-domain guided wavelet diffusion, belongs to the technical field of computer vision, and solves the problems that an existing diffusion model is high in calculation cost, easy to amplify noise and difficult to balance global and local details in low-illumination enhancement. Each of the training set and the test set comprises a low-illumination image and a reference image; decomposing the image into low-frequency and high-frequency components through discrete wavelet transform; only recovering a low-frequency component by using a diffusion model, introducing a detail weighted loss mechanism in training, and guiding low-frequency recovery by using high frequency; using the recovered low-frequency component as a priori, combining with a noisy high-frequency component, and performing feature modulation and denoising through a cross-frequency-domain gating refinement network; reconstructing a final enhanced image based on inverse wavelet transform; constructing a bidirectional cross-frequency-domain guided wavelet diffusion network model and designing a multi-objective loss function for training; and for the low-illumination image in the test set, the operation is executed based on the trained network model, and a high-fidelity enhanced image is obtained.

Description

A Low-Light Image Enhancement Method Based on Bidirectional Cross-Frequency Domain Guided Wavelet Diffusion Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, and specifically to a low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion. Background Technology

[0002] In computational photography and computer vision, low-light image enhancement has always been a fundamental and challenging task. Low-light images typically suffer from poor visibility, severe noise interference, loss of color information, and difficulty in discerning texture details. These degradation phenomena not only affect human visual perception but also severely limit the performance of downstream visual tasks such as object detection and semantic segmentation.

[0003] Early augmentation methods, such as histogram equalization and traditional Retinex theory, while providing solutions, often relied on manually designed priors, leading to artifacts or color distortion in the augmentation results. With the development of deep learning, methods based on convolutional neural networks (CNNs) have achieved end-to-end mapping, but their performance is highly dependent on the quality of large-scale paired data.

[0004] In recent years, generative models, especially denoising diffusion probabilistic models, have shown remarkable potential in image generation and inpainting. However, directly applying diffusion models to low-light enhancement tasks faces significant challenges: the iterative sampling process of diffusion models is computationally intensive, making it difficult to meet real-time requirements.

[0005] Applying diffusion models directly in the pixel domain often struggles to balance global illumination restoration with local detail preservation. The models tend to amplify noise indiscriminately while enhancing brightness, or generate fake textures inconsistent with the scene structure. Existing frequency-domain-based diffusion methods typically process each frequency band independently, ignoring the inherent connection between low frequencies (structure) and high frequencies (details), resulting in a lack of coherence in the generated edges and textures.

[0006] Therefore, how to design an enhancement method that can effectively reduce computational complexity and achieve high-fidelity detail restoration by coordinating low-frequency and high-frequency information is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] This invention addresses the problems of high computational cost, easy amplification of noise, and difficulty in balancing global structure and local details in existing diffusion models for low-light image enhancement tasks. It provides a low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion.

[0008] The technical solution adopted in this invention is:

[0009] A low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion includes the following steps:

[0010] Step S1: Obtain the training set and the test set respectively. Both the training set and the test set include low-light images and their corresponding normal-light reference images.

[0011] Step S2: Decompose the image into low-frequency and high-frequency components using K-level discrete wavelet transform;

[0012] Step S3: Use the diffusion model to recover only the low-frequency components, and introduce a detail-weighted loss mechanism during training to use high-frequency information to guide the low-frequency recovery.

[0013] Step S4: Using the recovered low-frequency component as a guide, combined with the noisy high-frequency component, feature modulation and denoising are performed through a cross-frequency gated refinement network (CFGR).

[0014] Step S5: Reconstruct the final enhanced image based on inverse discrete wavelet transform.

[0015] The present invention has the following beneficial effects:

[0016] Bidirectional cross-frequency domain guidance mechanism: This invention establishes a synergistic relationship between low and high frequencies. On the one hand, high-frequency energy is used to guide low-frequency diffusion training (detail-weighted loss) to ensure the accuracy of the structural basis; on the other hand, well-recovered low-frequency structure is used to guide the generation of high-frequency details (CFGR) to ensure the coherence of texture generation.

[0017] Efficiency and stability: By confining the diffusion process to compact low-frequency components, computational complexity and FLOPs are significantly reduced, while instability caused by direct diffusion sampling in high-frequency noise regions is avoided.

[0018] Excellent detail preservation and noise reduction capabilities: The designed cross-frequency domain gated refinement module combines guided residual enhancement and spatial gating mechanisms, which can adaptively suppress noise and sharpen edges. The generated images outperform existing mainstream methods in terms of SSIM and other metrics, and the visual effect is more natural. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0020] Figure 1 is a flowchart of the overall framework of bidirectional cross-frequency domain guided wavelet diffusion as described in Implementation Method 1.

[0021] Figure 2 is a module framework diagram of the bidirectional cross-frequency domain guided wavelet diffusion described in Implementation Method 1. Detailed Implementation

[0022] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] Implementation Method 1: The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion described in this implementation method, as shown in Figure 1, includes the following steps:

[0024] Step S1: Obtain the training set and test set respectively. Use a publicly available low-light image augmentation dataset, which contains pairs of low-light input images and normal-light reference images.

[0025] Step S2: Decompose the image using K-level discrete wavelet transform (K-DWT).

[0026] Specifically, let the input image be I, and the number of decomposition layers be K (K=2 is preferred in this embodiment). The decomposition yields a low-frequency component A (approximation coefficient) and a set of high-frequency components V, H, D (representing detail coefficients in the vertical, horizontal, and diagonal directions, respectively).

[0027] For low-light images Decomposition yields and high frequency components , , ;

[0028] For reference image Decomposition yields and high frequency components , , .

[0029] Since the low-frequency component size is only half the power of the original image, the diffusion model can significantly reduce the computational cost when operating in this space.

[0030] Step S3: Construct a detail-weighted diffusion model to recover low-frequency components.

[0031] Will As a conditional input, the diffusion model is trained to predict and remove noise to recover... To prevent the model from ignoring textured regions, a detail-weighted loss mechanism is introduced (see Implementation Method 2). This mechanism uses the high-frequency energy of the reference image to guide the loss calculation of the low-frequency components, prompting the model to focus on structural edges and obtain the recovered low-frequency components.

[0032] Step S4: Update the high-frequency components using the Cross-Frequency Gated Refinement Network (CFGR) model.

[0033] Due to the high frequency components of low-light images , , Containing significant noise and blurred details, it cannot be used directly. This step utilizes the high-quality low-frequency components recovered in step S3 as global prior guidance, which is input into the CFGR module. The CFGR module includes a Guided Residual Enhancement Module (GREM) and a Spatial Gated Refinement Module (SGRM) to perform feature modulation, denoising, and sharpening on the high-frequency components, outputting updated high-frequency components.

[0034] Step S5, image reconstruction.

[0035] The restored low-frequency components and the updated high-frequency components are then synthesized into the final enhanced image using K-order inverse discrete wavelet transform (K-IDWT). .

[0036] Steps S6 to S8 involve constructing a complete network model (BCF-Diff), designing a total loss function that includes weighted diffusion loss, reconstruction loss, and frequency loss, fine-tuning it end-to-end on the training set, and executing the above inference process on the test set.

[0037] This embodiment further defines the detail-weighted loss mechanism described in step S3 of embodiment one. Traditional diffusion model loss functions typically treat all pixels equally, resulting in smoothed texture details.

[0038] Implementation Method 2: This invention utilizes high-frequency components to construct a weighted graph. .

[0039] The specific construction method is as follows:

[0040] 1. Calculate the energy map of the high-frequency components of the reference image.

[0041] ;

[0042] This energy map reflects the distribution of edges and textures in the image.

[0043] 2. Upsampling and Normalization: Due to the low resolution of high-frequency components, bilinear interpolation is used to... Upsample to the same resolution as the low-frequency components or the original image, and normalize to the [0,1] interval to obtain .

[0044] 3. Generate the final weight map: Calculate the weights using a non-linear mapping function. :

[0045]

[0046] Here, a and b control the lower and upper limits of the weights, respectively (in this embodiment, a = 1.0, b = 5.0), and n is a hyperparameter controlling the shape of the curve (in this embodiment, n = 0.2). Since n < 1, the mapping curve is concave, which means that even weak energy fine texture regions can receive a significant weight boost, preventing them from being ignored.

[0047] 4. Apply weighted loss:

[0048]

[0049] in This represents element-wise multiplication. This forces the model to focus on regions with high-frequency energy concentrations during training.

[0050] Implementation Method 3: This implementation method is a further limitation on the Guided Residual Enhancement Module (GREM) described in step S4 of Implementation Method 1.

[0051] The guided residual enhancement module aims to establish a bridge between the recovered low-frequency components and the noisy high-frequency components. Since the low-frequency and high-frequency components have inherent spatial consistency, GREM does not perform pixel-level spatial alignment but instead employs a feature-level cross-frequency modulation strategy. GREM uses a dual-stream architecture, specifically including a guide stream and a content stream, and the specific processing procedure is as follows:

[0052] 1. Streaming and Modulation Parameter Generator (MPG)

[0053] The role of the pilot stream is to extract clean global context information from the recovered low-frequency components and convert it into modulation parameters.

[0054] First, the low-frequency components recovered in step S3 are input into the guiding stream, and projected onto a high-dimensional feature space through a 3×3 convolutional layer to obtain the guiding feature map. .

[0055] To capture the overall illumination distribution and structural overview of the image, the modulation parameter generator (MPG) performs... Perform global average pooling to compress spatial information into a compact global context descriptor z. The calculation formula is as follows:

[0056]

[0057] in, H and W are the height and width of the feature map.

[0058] The descriptor z is then fed into a lightweight network consisting of two 1×1 convolutional layers (with a LeakyReLU activation function in between). This network learns to map the global descriptor to scaling parameters for feature modulation. and bias parameters :

[0059]

[0060] in, and This is the weight matrix of the convolutional layer. (Generated) It was then reshaped and segmented to correspond to each Bootstrap Enhancement Block (GEB) in the subsequent content stream.

[0061] 2. Content Flow and Bootstrap Enhancement Block (GEB)

[0062] The purpose of content streaming is to use generated parameters to correct and enhance noisy high-frequency details.

[0063] The high-frequency components to be updated (containing three sub-bands: V, H, and D) are input into the content stream and projected into the same embedding space as the guiding features through a 3×3 convolutional layer to obtain the initial content features.

[0064] Content features are sequentially passed through k cascaded guided enhancement blocks (GEBs). The GEB is the core unit of GREM, and its internal operations include affine transformation modulation, nonlinear activation, and residual connection.

[0065] In the i-th GEB, the corresponding parameters generated by the guiding flow are used to... and Features of the input content Perform a channel-based affine transformation. This operation injects pure low-frequency global priors into the high-frequency features:

[0066]

[0067] in, This represents element-wise multiplication. In this way, The intensity of the feature was recalibrated, and This provides a reference offset, which adaptively enhances or suppresses specific high-frequency details based on the global lighting environment.

[0068] Modulated features Then, further spatial feature extraction is performed using the LeakyReLU activation function and a 3×3 convolutional layer. To maintain smooth gradient propagation and prevent information loss, GEB introduces an additive residual connection between the input and output.

[0069] Through the above mechanism, GREM effectively utilizes the recovered structural information (such as object contours and illumination distribution) in low-frequency components to guide the denoising of high-frequency components, ensuring that the generated texture details are consistent with the overall image environment and avoiding artifacts and clutter commonly found in traditional methods.

[0070] Implementation Method 4: This implementation method is a further limitation on the Spatial Gating Refinement Module (SGRM) described in step S4 of Implementation Method 1.

[0071] Although GREM restores the structure, random noise may still remain in the high-frequency components. SGRM is designed to filter out noise through a lightweight gating mechanism.

[0072] SGRM primarily comprises Subband Filtering Gate (SFG) and Simplified Channel Attention (SCA). Input features are first processed through LayerNorm and a 1×1 convolution, then through a 3×3 depthwise convolution (DWConv) to extract local spatial context. The features are then divided into two equal parts along the channel dimension. One part is used as a gate, activated by a sigmoid function, and then element-wise multiplied with the other part. This operation is similar to a dynamic filter, automatically suppressing low-response noise regions while preserving high-response edge details based on local features.

[0073] The filtered features are then weighted by channel to further highlight important feature channels.

[0074] The final output is added to the input via a residual connection to obtain a clean and sharpened high-frequency component.

[0075] Implementation Method 5: This implementation method is a further limitation on the loss function described in step S7 of Implementation Method 1.

[0076] To train the entire BCF-Diff network, this invention designs a total loss function:

[0077]

[0078] in, That is, the detail-weighted noise prediction loss described in Implementation Method 2.

[0079] Image reconstruction loss. Calculate the final enhanced image. Compared with reference image L1 distance and Structural Similarity (SSIM) loss between them:

[0080]

[0081] Frequency component loss. Direct supervision of the high-frequency components of the CFGR output, including L2 loss and total variational (TV) loss, is applied to suppress high-frequency artifacts.

[0082]

[0083] in , .

[0084] The above provides a detailed description of a low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion, characterized in that, Includes the following steps: Step S1: Obtain training and testing sets, each including low-light images and their corresponding normal-light reference images. Step S2: Decompose the images in the training set into low-frequency and high-frequency components using a K-order discrete wavelet transform. The high-frequency components include sub-bands in the vertical, horizontal, and diagonal directions. Step S3: Recover the low-frequency components of the low-light images using a diffusion model. During training, calculate the energy map of the high-frequency components to construct a detail-weighted loss mechanism, guiding the diffusion model to prioritize the recovery of structural details, thus obtaining the recovered low-frequency components. Step S4: Using the recovered low-frequency components as guiding information, combine them with the high-frequency components of the low-light images to perform feature extraction based on a cross-frequency gated refinement network model (CFGR). Modulation and denoising, updating high-frequency components; Step S5, the recovered low-frequency components and the updated high-frequency components are used for image reconstruction based on inverse discrete wavelet transform to obtain the enhanced image; Step S6, a bidirectional cross-frequency domain guided wavelet diffusion network model (BCF-Diff) is constructed based on the diffusion model and the cross-frequency domain gated thinning network model; Step S7, a total loss function is designed based on noise prediction weighted loss, image reconstruction loss and frequency component loss, and the bidirectional cross-frequency domain guided wavelet diffusion network model is trained using the total loss function; Step S8, for low-light images in the test set, based on the trained bidirectional cross-frequency domain guided wavelet diffusion network model, the operations of steps S2 to S5 are performed to obtain the final enhanced image.

2. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 1, characterized in that, In step S3, the construction method of the detail weighted loss mechanism is as follows: First, the detail energy map is obtained by calculating the sum of squares of the high-frequency components of the reference image in the vertical, horizontal and diagonal directions; then, the detail energy map is upsampled to the original image resolution and normalized by bilinear interpolation; finally, the normalized energy map is converted into the final weight map by a nonlinear mapping function, and the weight map is used to spatially weight the noise prediction loss during diffusion model training.

3. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 2, characterized in that, A nonlinear mapping function is used: the lower and upper limits of the weights are set respectively, and the shape of the curve is controlled by hyperparameters to make the curve concave to increase the weight of weak texture regions.

4. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 1, characterized in that, In step S4, the Cross-Frequency Gated Refinement Network Model (CFGR) consists of a Guided Residual Enhancement Module (GREM) and a Spatial Gated Refinement Module (SGRM) connected in series. The GREM receives the recovered low-frequency components and noisy high-frequency components, and uses the global statistical information of the low-frequency components as a priori to perform feature modulation on the high-frequency components. The SGRM receives the output of the GREM and suppresses residual noise and sharpens edges through gating and attention mechanisms.

5. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 4, characterized in that, The Guided Residual Enhancement Module (GREM) adopts a dual-stream architecture, including a guide stream and a content stream. The guide stream receives the recovered low-frequency components, projects them through the Conv module, and generates global modulation parameters through the Modulation Parameter Generator (MPG). The content stream receives the high-frequency components to be enhanced, projects them through the Conv module, processes them through multiple cascaded guided enhancement blocks (GEBs), and finally obtains the final output by adding the residuals and connecting them to the input.

6. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 5, characterized in that, The modulation parameter generator (MPG) first aggregates spatial information using global average pooling (GAP) to obtain a compact global context descriptor; then it generates corresponding scaling parameters γ and bias parameters β through two 1×1 convolutional layers; the γ and β are reshaped and segmented to correspond to each bootstrap enhancement block.

7. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 5, characterized in that, The Guided Enhancement Block (GEB) is as follows: first, the content stream features are modulated, and then processed by the Conv module and activation function; the feature modulation is achieved through an affine transformation formula.

8. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 4, characterized in that, The Spatial Gated Refinement Module (SGRM) is as follows: First, the input features are processed by Layer Normalization (LayerNorm) and 1×1 convolution. Then, the local spatial context is extracted by 3×3 depth convolution. Next, the features are enhanced by Subband Filtering (SFG) and Simplified Channel Attention (SCA). Finally, the final output is obtained by concatenating the summed residuals with the input through 1×1 convolution.

9. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 8, characterized in that, The subband filter gate (SFG) is as follows: the input features are split into two halves along the channel dimension; one half of the features is used as a dynamic filter, and the other half of the features are modulated by element-wise multiplication to suppress low-frequency noise and preserve high-frequency details.

10. The low-light image enhancement method based on bidirectional cross-frequency domain guided wavelet diffusion according to claim 1, characterized in that, In step S7, the total loss function is defined as: ;in, The noise prediction weighted loss of the diffusion model is calculated using the weight graph described in step S3; To enhance the combination of L1 loss and structural similarity (SSIM) loss between the image and the reference image; This is a combination of the L2 loss and the total variation (TV) loss between the enhanced high-frequency component and the reference high-frequency component.

Images