Mine low-light image enhancement denoising method and system based on wavelet diffusion model

Abstract

The invention discloses a mine low-light image enhancement denoising method and system based on a wavelet diffusion model, and the method comprises the steps: an initial state mapping and condition embedding module which is used for splicing an input low-light image and a noise distribution diagram, mapping the spliced low-light image and noise distribution diagram to a high-dimensional hidden space, and outputting a high-dimensional feature tensor; the time-frequency sensing and feature decoupling module is used for realizing time domain control and frequency domain separation; the double-current parallel adaptive enhancement module is used for processing the low-frequency characteristic component and the high-frequency characteristic component; the self-adaptive feature alignment module is used for mapping low-frequency and high-frequency feature values to a unified distribution reference, dynamically adjusting a scaling gain and eliminating reconstruction noise; and the dynamic reconstruction and imaging output module is used for realizing dynamic task scheduling, space restoration and artifact elimination after restoration. According to the method, dual optimization of enhanced quality and operation efficiency can be realized, non-uniform noise is accurately suppressed, illumination is balanced, and the problems of out-of-control image brightness, color cast, insufficient reliability and the like are solved on the premise of not depending on pairwise data.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a method and system for enhancing and denoising low-light images in mines based on a wavelet diffusion model. Background Technology

[0002] In the field of mine safety monitoring and intelligent operation, low-light image enhancement technology is a key means to improve the usability of low-quality monitoring images. It aims to improve the brightness and contrast of images captured in extremely low-light environments while suppressing noise. Currently, for image enhancement processing in complex low-light environments such as mines and underground roadways, existing technologies mainly focus on three directions: traditional image enhancement methods, deep learning-based low-light enhancement methods, and generative model-based image inpainting and enhancement methods. Traditional methods, such as histogram equalization, gamma correction, and multi-scale Retinex, have been applied in some mine monitoring systems. They mainly improve visual effects by making global or local adjustments to image brightness or contrast. With the development of deep learning technology, some research and products have begun to use convolutional neural networks for end-to-end enhancement of low-light images. By learning the mapping relationship between low-light images and normal-light images, they improve brightness and contrast. These methods have achieved certain results in natural scenes, but they still rely on a large amount of training data in mine environments and have limited adaptability to dust noise and extreme lighting distributions. Existing diffusion-based enhancement methods typically employ iterative denoising strategies, performing multiple rounds of evolution on low-light images in the spatial domain to recover high-quality, sharp images. To improve processing efficiency, some methods introduce hierarchical structures or subspace transformations, attempting to preserve image texture while reducing computational complexity.

[0003] While existing technologies have improved the visibility of low-light images to some extent, they still have significant shortcomings in complex industrial scenarios such as mines. First, there is the blindness in noise distribution processing. Existing diffusion models typically assume that noise is spatially uniformly distributed (e.g., standard Gaussian noise). However, noise in low-light images exhibits strong "signal dependence" and "spatial non-uniformity," meaning that noise is dense in dark areas while details are relatively good in bright areas. Traditional diffusion models perform uniform denoising on the entire image, leading to color artifacts easily appearing in dark areas, while textures in relatively bright areas are over-smoothed. Second, feature confusion caused by frequency component coupling. Most diffusion models directly extract and downsample features in the spatial domain, causing low-frequency information representing global brightness to intertwine with high-frequency information representing noise and texture. During denoising iterations, the model struggles to simultaneously and accurately complete illumination compensation (for low frequencies) and detail restoration (for high frequencies), often resulting in uneven brightness recovery or blurred edges. Third, the iterative process lacks adaptive guidance. The reverse denoising process of existing diffusion models often uses fixed computational logic at all time steps, failing to consider the different denoising timing requirements for different brightness inputs. For example, for extremely dark images, the model's need for supplemental lighting in the early stages is far greater than its need for detail enhancement in the later stages. However, the existing unified scheduling mechanism leads to unreasonable allocation of computing resources, increases the randomness of the generation process, and limits the stability of the enhancement effect. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for enhancing and denoising low-light images in mines based on a wavelet diffusion model. This method can achieve dual optimization of enhancement quality and computational efficiency, accurately suppress non-uniform noise and balance illumination, and solve the problems of uncontrolled image brightness, color cast and insufficient reliability without relying on paired data.

[0005] To achieve the above objectives, this invention provides a low-light image enhancement and denoising system for mines based on a wavelet diffusion model, comprising: The initial state mapping and conditional embedding module, including a channel stitching unit and a high-dimensional feature extractor, is used to process the input low-light image. Noise distribution map at the current time step Physical concatenation is performed along the channel dimension, and then mapped to a high-dimensional latent space using a projector, ultimately outputting a high-dimensional feature tensor containing information about the original texture structure. ; The time-frequency sensing and feature decoupling module is connected to the initial state mapping and conditional embedding module. It includes a time-sensing encoder (TE) and a discrete wavelet transform unit (DWT) to realize time-domain control and frequency-domain separation. The TE includes a sinusoidal position encoder and a multilayer perceptron (MLP). The dual-stream parallel adaptive enhancement module, connected to the time-frequency sensing and feature decoupling module, includes the adaptive low-frequency enhancement module LFEM based on global-local dual paths and the illumination-guided non-uniform denoising module HFEM, which are used to process the decoupled low-frequency and high-frequency feature components. LFEM integrates a global illumination estimator, a local texture extractor ResBlock, and a spatial gated fusion unit SGM, while HFEM integrates a guided feature extractor and a spatial feature transformer SFT. The adaptive feature alignment module, connected to the dual-stream parallel adaptive enhancement module, includes an instance normalization unit (IN), a local dark area analyzer, a dynamic gain predictor, and a residual scaling unit. It is used to map low-frequency and high-frequency feature values ​​to a unified distribution benchmark and dynamically adjust the scaling gain according to the current iteration progress to eliminate reconstruction noise caused by magnitude mismatch. The dynamic reconstruction and imaging output module, connected to the adaptive feature alignment module, includes a dynamic weight scheduler (PFS), an inverse wavelet transform unit (IDWT), and an anti-aliasing stitcher, used to achieve dynamic task scheduling, spatial reconstruction, and artifact elimination after reconstruction.

[0006] As a further aspect of the present invention: the time-frequency sensing and feature decoupling module inputs a high-dimensional feature vector. and the current time step scalar TE will scalar Parsing into a global modulation vector This is sent as a command signal to the subsequent enhancement module; the DWT unit utilizes the invertibility of wavelets to... Non-destructive disassembly into low-frequency subbands characterizing illumination / profile and high-frequency subbands characterizing noise / edges The final output consists of physically separated low-frequency and high-frequency feature components, as well as a global modulation vector. .

[0007] As a further aspect of the present invention: the low-frequency subband after input separation of the dual-stream parallel adaptive enhancement module. High-frequency subband and global modulation vector LFEM performs parallel computation of global brightness enhancement parameters and local detail residuals, and fuses them using an "illuminance confidence mask" generated by SGM to achieve brightness reshaping without overexposure. Then, the enhanced low-frequency features are used as illuminance priors to generate spatial modulation parameters. HFEM uses these spatial modulation parameters to perform pixel-level operations on the high-frequency subbands: strongly suppressing noise in dark areas and preserving edges in bright areas, ultimately outputting low-frequency features with balanced illumination. With high-frequency characteristics of precise noise reduction .

[0008] As a further aspect of the present invention: the adaptive feature alignment module inputs enhanced low-frequency features High-frequency characteristics Global modulation vector Local dark area intensity (LDI), IN is used to remove the mean and variance shifts of the dual-path features. The dynamic gain predictor is implemented using an MLP, based on... Alignment coefficient with LDI prediction The residual scaling unit is used to perform feature recalibration and residual concatenation, and finally outputs aligned features. and .

[0009] As a further aspect of the present invention: dynamic reconstruction and alignment of the dual-frequency features and LDI after imaging output module input, PFS based on LDI and current progress. Calculate the adaptive threshold and fusion weight In the early stages of iteration, lower-frequency flows are assigned higher weights, while higher-frequency flows are assigned higher weights in later stages. The IDWT unit losslessly restores the weighted fused frequency domain features to spatial domain features. The anti-aliasing stitcher is used to eliminate the mesh effect caused by block processing, and finally outputs the denoised prediction map of the current iteration step. Denoising prediction image Enhanced image at the end of the reasoning This is the final enhanced image.

[0010] To achieve the above-mentioned objectives, this invention also provides a denoising method based on the wavelet diffusion model-based low-light image enhancement and denoising system for mines, comprising the following steps: Step 1, Initial State Space Construction: The system generates a pure noise tensor with the same size as the original image, which follows a standard normal distribution. At this point, the image visually appears as random snowflake noise across the entire screen, without any semantic information, and is defined as a noise distribution map. ; Step 2, Reverse denoising iterative loop: The system starts from... arrive The countdown loop, at each time step Within the image data, the following steps are performed: Step 2.1, Conditional Anchoring and Projection: Project the current noise distribution map Compared with the original low-light image By concatenating and convolving, the contour and edge information of the original low-light image is injected into the front noise distribution map, and the original low-light image is transformed from pure random noise into a feature map with structured noise attached to the object contour distribution. Step 2.2, Frequency Domain Physical Separation: Discrete wavelet decomposition is performed on the feature map, which is torn into two parts of physical properties. The low-frequency component becomes a blurry energy distribution map, reflecting the current brightness level; the high-frequency component becomes an edge map with a black background, mixed with dense dust noise signals. Step 2.3, Illumination reshaping of low-frequency components: global branch calculates affine parameters to increase grayscale value, local branch repairs contrast, SGM module generates "anti-overexposure mask" for suppression, grayscale histogram shifts to the right to become brighter, but the center of the bright light source is protected by the mask, and low-frequency components change from dark and blurry to bright and uniformly illuminated blurry image. Step 2.4, Illumination-guided purification of high-frequency components: The system uses bright low-frequency components to generate a spatial mask. In the dark area indicated by the low-frequency components, the value of the high-frequency components is multiplied by a coefficient close to 0; in the bright area / edge area, it is multiplied by a coefficient close to 1 to achieve spatial non-uniform filtering. The noise signal corresponding to the dark background in the high-frequency components is erased, while the signal corresponding to the edge of the object is completely preserved. Step 2.5, Dynamic Weighting and Spatial Restoration: The system according to... The process involves superimposing the purified high-frequency components and the reshaped low-frequency components according to a certain rule, performing an inverse wavelet transform (IDWT), and then passing them through an anti-aliasing convolutional layer. The frequency domain information is then reconstructed back into the pixel space, eliminating the block effect caused by stitching, and generating a denoised prediction image for the current moment. ; Step 2.6, State Update: Using the diffusion sampling formula, combined with And random perturbations, calculate the image at the next time step. This will advance the noise reduction process and prepare for the next round of more refined restoration. Step 3, final image output: when At the end of the loop, the system outputs the final tensor. The image has undergone a complete evolution from "pure noise" to "noisy contours" and then to "clear image", achieving effective visual enhancement.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: First, a frequency decoupling module based on discrete wavelet transform was constructed, achieving high-fidelity dimensionality reduction and reconstruction of features. By utilizing the mathematically complete reversibility of wavelet transform to replace traditional stride convolution or pooling operations, this invention reduces the computational dimensionality of intermediate feature maps in the diffusion model while preserving the subtle textures and spatial structure information of the image to the maximum extent. This effectively solves the problem of irreversible information loss during downsampling in deep networks and improves the structural integrity of the enhancement results.

[0012] Second, an illumination-guided high-frequency non-uniform denoising module was constructed, achieving precise suppression of complex noise. By extracting the enhanced low-frequency illumination map as a spatial prior, a "perception-driven" denoising mechanism was implemented using spatial feature transformation, solving the problem of color spot artifacts that are prone to occur in extremely dark areas, while avoiding excessive smoothing of details in bright areas, and significantly improving the peak signal-to-noise ratio of the image.

[0013] Third, a low-frequency enhancement module combining global-local dual-branch and gating mechanisms was constructed to achieve natural and stable illumination compensation. By using a multilayer perceptron to learn global affine logic and combining it with residual local compensation, the model can adaptively adjust the gain according to the extremely uneven local illumination in the mine. This significantly improves the brightness of dark areas while effectively preventing overexposure, ensuring the color saturation and visual naturalness of the enhanced image.

[0014] Fourth, a dynamic weight scheduling mechanism based on the intensity of local dark areas was constructed, enabling refined task allocation during the iterative process. An adaptive threshold was set by calculating the duty cycle of local dark areas, and a modified inverse cosine annealing strategy was adopted, allowing the model to focus on macroscopic illumination restoration in the early stages and microscopic texture optimization in the later stages. This dynamic scheduling mode improves the semantic coherence of the generated images and enhances the model's robustness to inputs of varying darkness levels in the mine.

[0015] Fifth, a multi-dimensional unsupervised joint loss system for local perception was constructed, achieving high-precision enhancement free from dependence on pairwise data. By introducing gradient-guided structural anchoring loss, local patch exposure loss, and wavelet domain multi-scale regularization terms, this invention can automatically achieve brightness equalization, noise reduction, and color correction without the need for a truth reference image, significantly improving the algorithm's transferability and engineering application value in unknown mining scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a block diagram illustrating the system principle structure of the present invention; Figure 3 This is a diagram of the WADE-Net network structure of the present invention; Figure 4 This is a schematic diagram of the dual-stream parallel adaptive enhancement module of the present invention; Figure 5 This is a schematic diagram illustrating the enhanced effect of the present invention; Detailed Implementation

[0017] The present invention will be further illustrated by the following examples.

[0018] like Figure 2 As shown, a low-light image enhancement and denoising system for mines based on a wavelet diffusion model includes: The initial state mapping and conditional embedding module, including a channel stitching unit and a high-dimensional feature extractor, is used to process the input low-light image. Noise distribution map at the current time step Physical concatenation is performed along the channel dimension, and then mapped to a high-dimensional latent space using a projector, ultimately outputting a high-dimensional feature tensor containing information about the original texture structure. ; The time-frequency sensing and feature decoupling module is connected to the initial state mapping and conditional embedding module. It includes a time-sensing encoder (TE) and a discrete wavelet transform unit (DWT) to realize time-domain control and frequency-domain separation. The TE includes a sinusoidal position encoder and a multilayer perceptron (MLP). Specifically, the time-frequency sensing and feature decoupling module is input with a high-dimensional feature vector. and the current time step scalar TE will scalar Parsing into a global modulation vector This is sent as a command signal to the subsequent enhancement module; the DWT unit utilizes the invertibility of wavelets to... Non-destructive disassembly into low-frequency subbands characterizing illumination / profile and high-frequency subbands characterizing noise / edges The final output consists of physically separated low-frequency and high-frequency feature components, as well as a global modulation vector. .

[0019] A dual-stream parallel adaptive enhancement module is connected to the time-frequency sensing and feature decoupling module, such as... Figure 4 As shown, it includes an adaptive low-frequency enhancement module LFEM based on global-local dual paths and an illumination-guided non-uniform denoising module HFEM, which are used to process the decoupled low-frequency feature components and high-frequency feature components. LFEM integrates a global illumination estimator, a local texture extractor ResBlock, and a spatial gated fusion unit SGM, while HFEM integrates a guided feature extractor and a spatial feature transformer SFT. Specifically, the low-frequency subband after input separation of the dual-stream parallel adaptive enhancement module. High-frequency subband and global modulation vector LFEM performs parallel computation of global brightness enhancement parameters and local detail residuals, and fuses them using an "illuminance confidence mask" generated by SGM to achieve brightness reshaping without overexposure. Then, the enhanced low-frequency features are used as illuminance priors to generate spatial modulation parameters. HFEM uses these spatial modulation parameters to perform pixel-level operations on the high-frequency subbands: strongly suppressing noise in dark areas and preserving edges in bright areas, ultimately outputting low-frequency features with balanced illumination. With high-frequency characteristics of precise noise reduction .

[0020] The adaptive feature alignment module, connected to the dual-stream parallel adaptive enhancement module, includes an instance normalization unit (IN), a local dark area analyzer, a dynamic gain predictor, and a residual scaling unit. It is used to map low-frequency and high-frequency feature values ​​to a unified distribution benchmark and dynamically adjust the scaling gain according to the current iteration progress to eliminate reconstruction noise caused by magnitude mismatch. Specifically, the adaptive feature alignment module inputs enhanced low-frequency features. High-frequency characteristics Global modulation vector Local dark area intensity (LDI), IN is used to remove the mean and variance shifts of the dual-path features. The dynamic gain predictor is implemented using an MLP, based on... Alignment coefficient with LDI prediction The residual scaling unit is used to perform feature recalibration and residual concatenation, and finally outputs aligned features. and .

[0021] The dynamic reconstruction and imaging output module, connected to the adaptive feature alignment module, includes a dynamic weight scheduler (PFS), an inverse wavelet transform unit (IDWT), and an anti-aliasing stitcher, used to achieve dynamic task scheduling, spatial reconstruction, and artifact elimination after reconstruction.

[0022] Specifically, the dynamic reconstruction and imaging output module input-aligned dual-frequency features and LDI, PFS based on LDI and current progress. Calculate the adaptive threshold and fusion weight In the early stages of iteration, lower-frequency flows are assigned higher weights, while higher-frequency flows are assigned higher weights in later stages. The IDWT unit losslessly restores the weighted fused frequency domain features to spatial domain features. The anti-aliasing stitcher is used to eliminate the mesh effect caused by block processing, and finally outputs the denoised prediction map of the current iteration step. Denoising prediction image Enhanced image at the end of the reasoning This is the final enhanced image.

[0023] This invention constructs a dual-path enhancement system with time-domain awareness and illumination adaptation capabilities by deeply integrating the frequency-domain decoupling characteristics of the Discrete Wavelet Transform (DWT) unit into the inverse iterative framework of the conditional diffusion model. In engineering implementation, an initial projection layer is first introduced at each time step of the inverse denoising iteration to conditionally fuse the current noise feature map with the original low-light reference map and project it into the latent space. Subsequently, the features are decoupled into low-frequency sub-band features representing the subject contour and illumination, and high-frequency sub-band features representing edge noise, using the Discrete Wavelet Transform operator. To achieve dynamic intervention of time-domain information, the system constructs a time-step information injection (TE) mechanism. By generating a high-dimensional time embedding vector, it injects it into the main network using an affine transformation mechanism, thereby achieving micro-adjustment of the enhancement intensity at different denoising stages at the operator level, ensuring deep coupling between frequency-domain processing and the generative architecture.

[0024] In the decoupled enhancement logic, this invention adopts a collaborative processing scheme of "low frequency guiding high frequency". The low-frequency enhancement branch extracts brightness statistics through global average pooling and uses a multilayer perceptron to predict the global brightness scaling factor and offset to achieve preliminary linear illumination compensation. Subsequently, it refines local contrast through residual convolutional blocks and fuses the original features through a gating mechanism to prevent overexposure. The high-frequency enhancement branch utilizes the enhanced features output by the low-frequency branch and applies them to the high-frequency sub-band through spatial feature transformation. This allows the algorithm to dynamically adjust the denoising intensity according to the local illumination intensity, enhancing the denoising gain in areas with weak illumination and significant noise, while preserving high-frequency details in areas with rich texture. In addition, the system introduces a dynamic weight scheduling mechanism based on the brightness of the input image. It calculates an adaptive time step threshold based on the local average brightness and dynamically allocates the focus of high and low frequency enhancement tasks before and after this threshold through a modified cosine annealing strategy, achieving a smooth evolution from macroscopic illumination restoration to microscopic texture optimization.

[0025] Implementation methods include: 1. Basic framework and initial feature projection based on the conditional diffusion model This invention employs a conditional diffusion probability model as the underlying paradigm for iterative evolution. By redefining the noise prediction function in each iteration of the diffusion model, it integrates WADE-Net (wavelet adaptive diffusion enhancement network), which possesses frequency domain decoupling capabilities. Figure 3 As shown, the core prediction operator is embedded in it, and iterative enhancement is performed within each time step of the back diffusion process to achieve the restoration and denoising of low-light images of mines.

[0026] 1) Initial projection of input features At each time step of the reverse iteration The network first transforms the multi-source inputs into high-dimensional latent space features through an initial projection layer. To fully utilize the original low-light information and prevent structural drift during the enhancement process, this invention employs channel-dimensional stitching to incorporate the original low-light image. As a guiding condition, specifically, the current noise distribution map to be processed is used. Compared with the original low-light image The images are then stitched together and subsequently projected back to the latent space dimension through a shallow convolutional layer. This approach ensures that the physical structure of the original image serves as a "spatial anchor" to guide the denoising process in each iteration. The initial feature mapping process is expressed as follows: ; in, This is a noise distribution map for the current time step; The input is a low-light image; It is a high-dimensional feature tensor representing the initial input features of the WADE-Net network at the current step; Represents convolution deal with; This represents a splicing operation.

[0027] 2) Original Image Reconstruction Estimation and Backsampling In each iteration, the embedded WADE-Net predicts the noise distribution for the current step. As the iterations progress, the prediction results exhibit a monotonous evolution from blurry to clear. By sharing the same unsupervised joint loss constraint across all time steps, a high temporal consistency in semantics and structure is ensured between the images generated at different stages. This is based on the noise predicted at the current step and the feature map at the current time. Real-time simulation of denoised prediction maps for clear images (Enhanced image at the end of the reasoning) This is the final enhanced image. The denoising prediction map represents the model's expected mapping of the final enhancement result at the current denoising stage, and its calculation formula is as follows: ; This estimate is then used to assist in calculating the characteristic state at the next time step. : ; in, The cumulative scheduling coefficient during the diffusion process; The noise distribution predicted by the network; This is a reparameterized random term introduced in backsampling.

[0028] 2. Time Step Information Injection Mechanism (TE) Because the denoising focus of the diffusion model differs at different stages (early stage focuses on macroscopic illumination restoration, later stage focuses on local detail enhancement), the network must be aware of the current time step. It also provides the ability to adjust the feature distribution in real time. This invention establishes a time-domain driven feature modulation mechanism to provide dynamic control signals for the subsequent dual-path enhancement module.

[0029] 1) High-dimensional feature encoding at time steps scalar time step Since deep networks cannot be directly input, they need to be mapped to high-dimensional vectors containing topological relationships. This invention utilizes a sinusoidal periodic function to generate a time position reference vector. The process is represented as follows: ; Subsequently, the baseline vector is projected through two fully connected layers and a nonlinear activation function to generate a temporally embedded global feature vector with strong semantic expressive power. : ; in, This is the time step number of the current reverse iteration; The dimension for encoding (usually 128 or 256); It is a baseline vector that contains temporal and positional characteristics; This is the encoded global modulation vector; Represents fully connected layer processing; This is the activation function.

[0030] 2) Feature modulation injection based on affine transformation To achieve microscopic control of enhancement intensity through temporal information, this invention employs adaptive feature modulation technology. It is injected into the core residual block of WADE-Net as a lateral branch. Then this mechanism... The gain coefficient (Scale) and bias coefficient (Shift) required for each convolutional layer are predicted in real time, and the modulated feature output formula is expressed as follows: ; in, This is the intermediate feature map after convolution processing of each branch, that is, the local image features in the current residual block that have undergone preliminary convolution processing but have not yet been temporally modulated; The output features in the formula are injected with time information, resulting in a feature map after distribution alignment. The group normalization operation is used to... Distribution standardization; and It consists of two independent fully connected layers from The scaling factor and bias factor are extracted and used to adjust the variance and mean of the features, respectively.

[0031] In this way, time step It directly controls the gain and offset of image features, especially in low-light scenes. When it is large, The brightness recovery intensity of the amplified low-frequency branch will be increased when... When decreasing, This "time-domain driven frequency-domain processing" design guides the high-frequency branch to perform refined texture compensation, ensuring that the model... The processing focus can be adaptively switched in each iteration.

[0032] 3. Dual-path enhancement processing mechanism based on wavelet frequency decoupling This invention integrates the Discrete Wavelet Transform (DWT) unit at the downsampling position of WADE-Net. By frequency decoupling, image features are transformed into low-frequency components with concentrated energy and high-frequency components with rich details. Targeted enhancement subnetworks are then constructed to achieve decoupled processing of illumination restoration and accurate noise reduction.

[0033] 1) Frequency space projection and decoupling The network first uses DWT to treat the initially projected feature map as a high-order feature tensor. Decomposed into low-frequency subbands and high-frequency subbands in three directions Compared to traditional stride convolution, DWT is mathematically lossless. It transfers spatial information to the frequency dimension instead of discarding it directly. This allows for physical separation of noise components (high frequencies) and contour components (low frequencies) while reducing spatial resolution, avoiding information loss common to traditional operators and laying the foundation for subsequent targeted enhancement. The process is represented as follows: ; in, This refers to the low-frequency subband after wavelet decomposition. These are the high-frequency subbands in different directions after wavelet decomposition; This represents wavelet transform using discrete wavelet transform units.

[0034] 2) Adaptive Low-Frequency Enhancement Module LFEM Based on Global-Local Dual Path Since the low-frequency subband after wavelet decomposition concentrates most of the energy and brightness information of the image, this invention constructs a dual-path architecture to handle the complex illumination distribution in the mine. The core function of the low-frequency enhancement path is to process the illumination distribution and global contour of the image, while significantly improving the brightness, avoiding contrast loss and local overexposure. It adopts a "global-local" dual-branch parallel processing architecture, combined with a gating fusion mechanism, to ensure the naturalness of illumination compensation.

[0035] Since improving the overall brightness of a low-light image requires considering the average illumination level of the entire image, global statistical feature extraction is performed first, utilizing the low-frequency subband. Brightness reshaping is performed based on global statistical characteristics to achieve global illumination compensation, and global average pooling (GAP) is used to extract... The macroscopic illuminance distribution is shown in the following formula: ; in, This is the compressed global feature vector; These are the height and width after wavelet decomposition; This is a global average pooling operation.

[0036] Next, a global brightness linear mapping is performed. Low-light images often appear generally dark and have low contrast, which cannot be balanced by simple addition. This branch simulates the affine transformation of physical illumination to achieve illumination enhancement across the entire image. The affine transformation parameter coefficients learned by the multilayer perceptron (MLP) can adaptively determine the supplementary light intensity according to the darkness of different images, achieving large-area illumination enhancement. The extracted global feature vector is input into an MLP to predict the scaling factor. and offset The process is represented as follows: ; ; in, To learn the affine transformation parameters of physical illumination through MLP, contrast stretching and brightness enhancement are controlled separately; This is the output sub-band of the global branch.

[0037] Global mapping can easily lead to blurred image details, so a dedicated branch is needed to extract local gradient information. Since residual structures can effectively capture the edges and local contrast of an image and compensate for the high-dimensional semantic information lost in global linear adjustment, residual connections are introduced to repair local contrast and prevent the loss of details caused by global adjustment.

[0038] In the initial stages of repairing the low-light environment in the mine, the TE module, through its larger derivatives... The gain coefficient significantly enhances the local branch's ability to capture low-frequency illumination gradients, assisting the global branch in rapidly increasing the energy of dark areas. In the later stages of iteration, this branch performs subtle contrast correction, ensuring the smoothness and consistency of the low-frequency background enhancement process through real-time feedback of temporal information. Specifically, it affects the input low-frequency sub-band. implement Convolution yields intermediate features, which are then fed into the TE module for utilization. Affine transformation is performed to obtain the modulated output. Then, it is used as input for residual connection to obtain the output of the local branch. The process is as follows: ; in, For the output sub-band of a local branch; The time step information injection mechanism is used to process the low-frequency subband. The output subband after that; is the activation function. .

[0039] In areas of extremely high brightness, the supplementary lighting weight should be reduced to prevent overexposure; in areas of extreme darkness, the supplementary lighting weight should be increased. Therefore, a spatial gating machine (SGM) container is designed for adaptive gating fusion. This fusion method achieves "illuminance perception," adjusting the weight allocation in areas requiring significant supplementary lighting, so that the final result achieves sufficient brightness while preserving solid local structure. Specifically, to smoothly stitch together the global and local enhancement results, a spatial gating mechanism is introduced to generate a weight mask. The adaptive fusion of illuminance is achieved, and the process is represented as follows: ; ; in, This is a weight matrix, where each value represents the degree of confidence in the supplemental lighting results at the current location. For Hadamard product, it means multiplying pixel by pixel; As an enhanced low-frequency subband, it is not only used for reconstruction, but also as a guide signal sent into the high-frequency path; represent Convolution processing; This is the activation function.

[0040] 3) Illumination-guided non-uniform noise reduction module HFEM To address the uneven noise distribution in mines (more noise in dark areas and finer texture in bright areas), HFEM utilizes... Using an "illuminance map" to guide high-frequency denoising, this module addresses the issues of illusions and color spots in extremely dark areas while preserving texture in bright regions. This module employs the Spatial Feature Transform (SFT) for implementation, and because the SFT operation occurs after wavelet decomposition... Since it is performed at the scale and only involves pixel-by-pixel arithmetic operations, its computational cost is much lower than that of global attention calculation at the original image scale, thus improving computational efficiency.

[0041] SFT passed The project achieves "spatial awareness," automatically shrinking in identified extremely dark areas to eliminate hallucinations in the random noise generated by the diffusion model. This method enables the model to "inject" the structural edges that have been recovered from the low-frequency components into the high-frequency path while suppressing noise, in order to solve the problem of blurred edges after low-light image enhancement.

[0042] First, it is necessary to extract deep features that reflect changes in spatial illumination from the low-frequency components. Using depthwise separable convolutions can greatly reduce computational overhead while maintaining the receptive field. Therefore, the enhanced low-frequency components are input into a lightweight feature extraction network consisting of two layers of depthwise separable convolutions to obtain the guided feature map. The process is represented as follows: ; in, This is the guiding feature map for the high-frequency enhancement module; represent Depth-separable convolution processing.

[0043] Then, based on the guiding features It learns the scaling factor and offset corresponding to each pixel position in space. Specifically, it does this through two parallel... The convolutional layer maps the scaling factor matrix and offset matrix from the generated guided feature map, as shown in the following formula: ; ; in, This is a scaling factor, representing the proportion of high-frequency signals retained. In areas with extremely low light, noise dominates. The value will tend to 0, thus suppressing noise in well-lit areas. The value tends towards 1, preserving the original texture; is the offset factor, representing the texture compensation term.

[0044] Residual information is extracted from the edges of the low-frequency structure to compensate for the weak high-frequency features that may be lost during the denoising process. This residual information is then applied to each high-frequency sub-band, and a linear affine transformation is performed on each high-frequency sub-band. Finally, the SFT-modulated high-frequency components are output. The spliced ​​high-frequency sub-bands are then modulated, as shown in the following formula: ; in, For Hadamard product, it means multiplying pixel by pixel; This refers to the high-frequency subband output after modulation. This refers to the high-frequency wavelet components after frequency decoupling.

[0045] Finally, in order to further correct the slight distortion that may be introduced during SFT modulation and enhance the edges, the following steps were taken: Perform residual refinement operations and simultaneously receive adjustments to the initial denoising. By performing time-domain driven detail optimization, the final high-frequency output is obtained. .

[0046] In the later stages of the diffusion model iteration, the TE module detects that the image is close to being sharp, and at this point, it adjusts... The bias term guides the model to inject more edge texture residuals into the high-frequency branches, effectively compensating for the weak high-frequency features that may be lost during the initial strong denoising process. This time-domain-aware high-frequency repair mechanism enables the model to significantly improve the contour sharpness of the target object while suppressing mine dust noise.

[0047] Specifically, for the input subband implement Convolution yields intermediate features, which are then fed into the TE module for utilization. Affine transformation is performed to obtain the modulated output. Then, it is used as input for residual connection to obtain the output of the local branch. The process is as follows: ; in, For high-frequency subband output by the high-frequency enhancement path; The time step information injection mechanism is used to process high-frequency subbands. The output sub-band after that.

[0048] 4. Dynamic weight scheduling and feature space reconstruction In the inverse denoising loop of the diffusion model, with time step... from As the value gradually decreases to 1, the focus of the model's processing will naturally transition from global illumination restoration to local detail refinement. This invention uses the dynamic weight scheduler PFS to control the enhancement focus in real time and combines feature alignment and spatial stitching techniques to achieve high-quality image reconstruction.

[0049] 1) Local dark area intensity sampling and adaptive threshold setting First, local illuminance sampling is performed. To overcome the interference of local point light sources such as miner's lamps in illuminance evaluation, this invention uses Local Dark Area Intensity (LDI) as the scheduling benchmark. The original low-light input image is first processed... Divided into Non-overlapping patches Calculate the average brightness of each patch and select the set of patches with the lowest brightness (50%). Calculate its average illuminance scalar The process is as follows: ; in, Average brightness for each patch; The average illuminance scalar value accurately reflects the extreme darkness of the area to be enhanced in the mine scene. Its value range has been normalized. ; This is a set of patches with the lowest brightness (50%). This is a function operation for calculating the average value.

[0050] Then, the critical time threshold is calculated. The dynamic determination diffusion model should focus on the time span of low-frequency supplemental lighting, when At that time, it was considered that the image was still in a stage of insufficient macroscopic illumination; when At this point, the image is considered to have entered the fine-tuning stage. Its formula is expressed as follows: ; in, The total number of iterations set for backdiffusion in the diffusion model; The threshold step number is obtained through adaptive calculation.

[0051] 2) Adaptive Feature Alignment (AFA) Since LFEM focuses on significant brightness enhancement, while HFEM focuses on minor noise filtering, the two output features may differ significantly in magnitude. Therefore, feature alignment is performed before fusion to ensure reconstruction stability. Aligned features not only eliminate reconstruction noise caused by magnitude mismatch but also... The introduction of this feature enables the alignment process to be intelligently adjusted according to the progress of the diffusion iteration.

[0052] First, instance normalization is performed. For and Instance normalization is performed separately, and mean and variance offsets are removed separately to map the two features to a unified distribution benchmark, thereby eliminating the influence of absolute energy differences.

[0053] Then residual scaling mapping is performed, utilizing the global modulation vector embedded at the current time step. Using the intensity of local dark areas as a modulating prior, a lightweight fully connected network is used to predict the scaling gain required for the dual-path features in the current denoising stage. and As shown in formula (23), the generated alignment coefficients are then used to perform dynamic gain compensation on the normalized features. In order to maintain the stability of the residual structure, this invention retains a certain proportion of the original feature information during the alignment process. Finally, the aligned low-frequency features High frequency characteristics It is expressed as follows: ; ; ; in, For instance normalization process; and These are the predicted low-frequency and high-frequency alignment coefficients, respectively.

[0054] 3) Dynamic weight allocation based on the modified cosine formula To avoid visual discontinuities caused by sudden weight changes, this invention employs a modified inverse cosine annealing formula to calculate the fusion weights at each step. In the early stages of iteration, the network heavily relies on the output from the low-frequency enhancement module. Ensure a rapid increase in illumination; in the later stages of iteration, the network focus shifts towards the output of the high-frequency enhancement module. Tilting is used to eliminate residual noise and sharpen edges. The weight calculation formula is shown below: ; ; in, Low-frequency weights; High-frequency weights; This is the current time step of the reverse iteration; The preset weights.

[0055] 4) Feature weighted fusion and inverse wavelet transform reconstruction The aligned frequency domain components are weighted and aggregated according to dynamic weights, and the feature map is restored to the spatial domain without loss through inverse wavelet transform IDWT. The process is as follows: ; ; in, and These are the input and output of IDWT, respectively. By using IDWT instead of traditional bilinear interpolation or deconvolution, IDWT can utilize the orthogonality of wavelet components to losslessly map the weighted frequency domain features back to the spatial domain feature map.

[0056] 5) Anti-serrated stitching and finishing output To address the potential mesh artifacts at frequency boundaries in wavelet transform, this invention introduces an anti-aliasing stitching layer after IDWT. This layer performs local stitching and smoothing of the reconstructed spatial features through cross-scale convolution. ; in, This is the final output of this iteration of the diffusion model.

[0057] The residual connections in Equation (28) ensure the propagation of the underlying structure, while the convolutional layers eliminate the mesh texture generated by reconstruction by sensing the transition between adjacent frequency blocks. The final output... As a prediction result, it is used to calculate the noise estimate for the current time step. This drives the diffusion model into the next sampling cycle.

[0058] 5. Unsupervised joint loss function To achieve self-consistent training for image enhancement and denoising under unsupervised conditions, this invention defines a joint loss function consisting of spatial structure preservation, wavelet domain feature regularization, and exposure color balance, ensuring the suppression of illusions and the achievement of high-quality enhancement without a reference image.

[0059] To achieve self-consistent training for mine image enhancement and denoising under unsupervised conditions, this invention defines a multi-dimensional joint loss function. This system estimates the enhanced image during the diffusion loop. Constraining the intermediate frequency domain features ensures that the generated results not only conform to the physical characteristics of the mine but also possess a high degree of structural integrity.

[0060] 1) Loss of spatial structural consistency To suppress the visual illusions produced by diffusion models in dark areas, this invention introduces a gradient-guided structural anchoring mechanism to forcibly enhance the image. Gradient distribution and original low-light image To maintain consistency, first extract the gradient using the Sobel gradient operator. and Edge information in the horizontal and vertical directions and Then, based on the input image at the pixel level... Local variance calculation structure anchor weights As shown in formula (30), the expression for this loss function is as follows: ; ; in, This is the Sobel gradient operator, used to extract edge information between pixels; and They are respectively and In pixels Gradient information at the location; This represents the total number of pixels. The input image in pixels Local variance at; For structural anchoring weights based on input local variance. In extremely dark regions, The signal is weak. Minimal, at this time Enlarge, Force Must be strictly followed The weak gradients suppress the random generation of illusion textures.

[0061] 2) Local adaptive exposure loss To address the extremely uneven lighting conditions in mines, this invention abandons global average brightness constraints and instead employs local patch-level exposure control to avoid localized overexposure or complete darkness. This approach forces each local area of ​​the image to fall within the brightness range most comfortable for human vision, achieving automatic illumination equalization. First, the enhanced image... Split into The size is Non-overlapping local patches Then, a preset ideal exposure threshold is set, and the Euclidean distance between the mean value and the ideal value of each patch is calculated. The expression for this loss function is as follows: ; in, The number of patches in the partition; Average exposure for each patch; This is the preset ideal exposure threshold.

[0062] 3) Wavelet domain multi-scale denoising regularization term By utilizing the sparsity principle of wavelet components, the intermediate frequency domain output in the enhancement branch is directly constrained to ensure the smoothness of illumination and the thoroughness of noise reduction. This loss function includes two terms: the first term is the high-frequency subband. Sparsity constraints are applied to the three refined high-frequency subbands, in conjunction with a non-uniform denoising matrix derived from SFT. The first term is to suppress noise energy in non-edge regions; the second term is the total variational regularization of the low-frequency components, which affects the enhanced low-frequency subband. A second-order total variation constraint is applied to ensure spatial smoothness of the brightened background. The final loss function expression is shown below: ; in, For hyperparameters; The non-uniform denoising matrix represents the denoising penalty strength, where... The scaling factor matrix from the SFT in the high-frequency enhancement module, through and The complementary relationship between these elements achieves mathematical unity between the positive processing of the enhancement module and the negative constraint of the loss function, ensuring that the model can suppress noise in extremely dark regions while simultaneously achieving optimal performance. This accurately compensates for the edge structure information lost due to noise reduction.

[0063] 4) Loss of color constancy Based on the "Grey World Theory," this approach addresses the color cast issues, such as reddish or bluish tints, that easily occur in mine scenes when brightness is significantly increased. This loss forces the energy ratio of the three channels to tend towards balance, ensuring the natural and realistic colors of the enhanced image. The enhanced image is calculated separately. Average pixel intensity in three channels Then, the squared differences between each pair of the three channels are calculated. The expression for this loss function is as follows: ; in, The average pixel intensity of the channel.

[0064] This invention constructs the final total loss function through weighted summation and simultaneously updates all learnable parameters of WADE-Net and related modules using the backpropagation algorithm. The expression for the total joint loss function is as follows: ; in, These are preset hyperparameters used to balance the contributions of each optimization objective.

[0065] This invention achieves a bright, clear, and natural-looking enhancement effect without relying on pairwise truth data. The specific implementation process is as follows: The first stage is data preparation and model training.

[0066] Step 1: Data selection. This invention is trained and validated on the LOL-v1 dataset, and three types of low-light images are used as the visualization of the test data. The results are as follows: Figure 5 The images shown are indoor, outdoor, and low-light images from a mine, respectively. Figure a shows the indoor low-light image and enhancement result; Figure b shows the outdoor low-light image and enhancement result; and Figure c shows the actual low-light image from a mine and enhancement result.

[0067] Step 2: Input the low-light data into the network and calculate the Local Dark Area Intensity (LDI) of each low-light image as a priori label. During training, calculate the designed joint loss function and iteratively update the network parameters through backpropagation until the loss function converges.

[0068] The second stage, such as Figure 1 The image enhancement process is shown in the diagram.

[0069] Step 1, Initial State Space Construction. The system generates a pure noise tensor with the same size as the original image, following a standard normal distribution. At this point, the image visually appears as random snowflake noise across the entire screen, without any semantic information, and is defined as a noise distribution map. ; Step 2, reverse denoising iterative loop. System startup from arrive The countdown loop, at each time step Within the image data, the following detailed processing steps are performed: Step 2.1, Conditional Anchoring and Projection. Project the current noise distribution map. Compared with the original low-light image The images are concatenated and convolved. Information such as the contours and edges of the original image is injected into the noise map. The image transforms from purely random noise into structured noise that is distributed according to the contours of the objects.

[0070] Step 2.2, Frequency Domain Physical Separation. Discrete wavelet decomposition is performed on the feature map, tearing the image into two parts of physical properties. The low-frequency component becomes a blurred energy distribution map, reflecting the current brightness level; the high-frequency component becomes a black-background edge map, mixed with dense dust noise signals.

[0071] Step 2.3, Illumination reshaping of low-frequency components. The global branch calculates affine parameters to increase grayscale values, the local branch repairs contrast, and the SGM module generates an "anti-overexposure mask" for suppression. The grayscale histogram shifts to the right, becoming brighter, but the center of the bright light source is protected by the mask. The low-frequency components transform from a dark and blurry image into a bright and uniformly illuminated blurry image.

[0072] Step 2.4, Illumination-guided purification of high-frequency components. The system uses bright low-frequency components to generate a spatial mask. In dark areas indicated by low-frequency components, the value of the high-frequency components is multiplied by a coefficient approaching 0; in bright / edge areas, it is multiplied by a coefficient approaching 1, achieving spatial non-uniform filtering. Noise signals in the high-frequency components corresponding to dark backgrounds are removed, while signals corresponding to object edges are completely preserved.

[0073] Step 2.5, Dynamic Weighting and Spatial Restoration. System According to the progress, the purified high-frequency components and the reshaped low-frequency components are superimposed according to a certain rule ratio, and the inverse wavelet transform (IDWT) is performed. Then, after passing through an anti-aliasing convolutional layer, the frequency domain information is reconstructed back into the pixel space, eliminating the block effect caused by stitching, and generating the prediction image for the current time step. Compared to the previous moment, its brightness has slightly increased, noise has slightly decreased, and the texture of objects is more realistic.

[0074] Step 2.6, State Update. Using the diffusion sampling formula, combined with... And random perturbations, calculate the image at the next time step. We are advancing the noise reduction process to prepare for the next round of more refined restoration.

[0075] Step 3, final image output. When At the end of the loop, the system outputs the final tensor. The image has undergone a complete evolution from "pure noise" to "noisy contours" and then to "sharp image". The final result achieves effective visual enhancement.

Claims

1. A mine low-light image enhancement and denoising system based on a wavelet diffusion model, characterized in that, include: The initial state mapping and conditional embedding module, including a channel stitching unit and a high-dimensional feature extractor, is used to process the input low-light image. Noise distribution map at the current time step Physical concatenation is performed along the channel dimension, and then mapped to a high-dimensional latent space using a projector, ultimately outputting a high-dimensional feature tensor containing information about the original texture structure. ; The time-frequency sensing and feature decoupling module is connected to the initial state mapping and conditional embedding module. It includes a time-sensing encoder (TE) and a discrete wavelet transform unit (DWT) to realize time-domain control and frequency-domain separation. The TE includes a sinusoidal position encoder and a multilayer perceptron (MLP). The dual-stream parallel adaptive enhancement module, connected to the time-frequency sensing and feature decoupling module, includes the adaptive low-frequency enhancement module LFEM based on global-local dual paths and the illumination-guided non-uniform denoising module HFEM, which are used to process the decoupled low-frequency and high-frequency feature components. LFEM integrates a global illumination estimator, a local texture extractor ResBlock, and a spatial gated fusion unit SGM, while HFEM integrates a guided feature extractor and a spatial feature transformer SFT. The adaptive feature alignment module, connected to the dual-stream parallel adaptive enhancement module, includes an instance normalization unit (IN), a local dark area analyzer, a dynamic gain predictor, and a residual scaling unit. It is used to map low-frequency and high-frequency feature values ​​to a unified distribution benchmark and dynamically adjust the scaling gain according to the current iteration progress to eliminate reconstruction noise caused by magnitude mismatch. The dynamic reconstruction and imaging output module, connected to the adaptive feature alignment module, includes a dynamic weight scheduler (PFS), an inverse wavelet transform unit (IDWT), and an anti-aliasing stitcher, used to achieve dynamic task scheduling, spatial reconstruction, and artifact elimination after reconstruction.

2. The mine low-light image enhancement and denoising system based on wavelet diffusion model according to claim 1, characterized in that, The time-frequency sensing and feature decoupling module takes a high-dimensional feature vector as input. and the current time step scalar TE will scalar Parsing into a global modulation vector This is sent as a command signal to the subsequent enhancement module; DWT units utilize the invertibility of wavelets to... Non-destructive disassembly into low-frequency subbands characterizing illumination / profile and high-frequency subbands characterizing noise / edges The final output consists of physically separated low-frequency and high-frequency feature components, as well as a global modulation vector. .

3. The mine low-light image enhancement and denoising system based on wavelet diffusion model according to claim 2, characterized in that, Low-frequency subband after input separation of the dual-stream parallel adaptive enhancement module High-frequency subband and global modulation vector LFEM performs parallel computation of global brightness enhancement parameters and local detail residuals, and fuses them using an "illuminance confidence mask" generated by SGM to achieve brightness reshaping without overexposure. Then, the enhanced low-frequency features are used as illuminance priors to generate spatial modulation parameters. HFEM uses these spatial modulation parameters to perform pixel-level operations on the high-frequency subbands: strongly suppressing noise in dark areas and preserving edges in bright areas, ultimately outputting low-frequency features with balanced illumination. With high-frequency characteristics of precise noise reduction .

4. The mine low-light image enhancement and denoising system based on wavelet diffusion model according to claim 3, characterized in that, The adaptive feature alignment module inputs enhanced low-frequency features High-frequency characteristics Global modulation vector Local dark area intensity (LDI), IN is used to remove the mean and variance shifts of the dual-path features. The dynamic gain predictor is implemented using an MLP, based on... Alignment coefficient with LDI prediction The residual scaling unit is used to perform feature recalibration and residual concatenation, and finally outputs aligned features. and .

5. The mine low-light image enhancement and denoising system based on wavelet diffusion model according to claim 4, characterized in that, Dynamic reconstruction and imaging output module input alignment of dual-frequency features and LDI, PFS based on LDI and current progress. Calculate the adaptive threshold and fusion weight In the early stages of iteration, lower frequency flows are given higher weights, while higher frequency flows are given higher weights in the later stages. The IDWT unit losslessly restores the weighted fused frequency domain features to spatial domain features; the anti-aliasing stitcher is used to eliminate the grid effect caused by block processing, and finally outputs the denoised prediction map of the current iteration step. Denoising prediction image Enhanced image at the end of the reasoning This is the final enhanced image.

6. A denoising method for a mine low-light image enhancement and denoising system based on a wavelet diffusion model as described in any one of claims 1-5, characterized in that, Includes the following steps: Step 1, Initial State Space Construction: The system generates a pure noise tensor with the same size as the original image, which follows a standard normal distribution. At this point, the image visually appears as random snowflake noise across the entire screen, without any semantic information, and is defined as a noise distribution map. ; Step 2, Reverse denoising iterative loop: The system starts from... arrive The countdown loop, at each time step Within the image data, the following steps are performed: Step 2.1, Conditional Anchoring and Projection: Project the current noise distribution map Compared with the original low-light image By concatenating and convolving, the contour and edge information of the original low-light image is injected into the front noise distribution map, and the original low-light image is transformed from pure random noise into a feature map with structured noise attached to the object contour distribution. Step 2.2, Frequency Domain Physical Separation: Discrete wavelet decomposition is performed on the feature map, which is torn into two parts of physical properties. The low-frequency component becomes a blurry energy distribution map, reflecting the current brightness level; the high-frequency component becomes an edge map with a black background, mixed with dense dust noise signals. Step 2.3, Illumination reshaping of low-frequency components: global branch calculates affine parameters to increase grayscale value, local branch repairs contrast, SGM module generates "anti-overexposure mask" for suppression, grayscale histogram shifts to the right to become brighter, but the center of the bright light source is protected by the mask, and low-frequency components change from dark and blurry to bright and uniformly illuminated blurry image. Step 2.4, Illumination-guided purification of high-frequency components: The system uses bright low-frequency components to generate a spatial mask. In the dark area indicated by the low-frequency components, the value of the high-frequency components is multiplied by a coefficient close to 0; in the bright area / edge area, it is multiplied by a coefficient close to 1 to achieve spatial non-uniform filtering. The noise signal corresponding to the dark background in the high-frequency components is erased, while the signal corresponding to the edge of the object is completely preserved. Step 2.5, Dynamic Weighting and Spatial Restoration: The system according to... The process involves superimposing the purified high-frequency components and the reshaped low-frequency components according to a certain rule, performing an inverse wavelet transform (IDWT), and then passing them through an anti-aliasing convolutional layer. The frequency domain information is then reconstructed back into the pixel space, eliminating the block effect caused by stitching, and generating a denoised prediction image for the current moment. ; Step 2.6, State Update: Using the diffusion sampling formula, combined with And random perturbations, calculate the image at the next time step. This will advance the noise reduction process and prepare for the next round of more refined restoration. Step 3, final image output: when At the end of the loop, the system outputs the final tensor. The image has undergone a complete evolution from "pure noise" to "noisy contours" and then to "clear image", achieving effective visual enhancement.

Images