Low-light image enhancement method and apparatus based on wavelet frequency domain residual quotient learning

Abstract

This application relates to a method and apparatus for low-light image enhancement based on wavelet frequency domain residual quotient learning. The method includes: performing a two-dimensional discrete wavelet transform on a preprocessed low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; inputting the low-frequency sub-bands into a low-frequency illumination prediction module to obtain a predicted illumination ratio map; obtaining enhanced low-frequency components based on the predicted illumination ratio map and the low-frequency sub-bands; inputting the three high-frequency detail sub-bands and the low-frequency sub-bands into a high-frequency suppression module to obtain corresponding enhanced high-frequency components; and inputting the enhanced low-frequency components and the three high-frequency components into an image reconstruction module to obtain an enhanced low-light image. This method embeds residual quotient learning into the wavelet domain, achieving interpretable enhancement in the frequency dimension. This method inherits the physical consistency of residual quotient learning while introducing the sparse representation capability of the wavelet domain, providing a new perspective for low-light enhancement tasks.

Description

Technical Field

[0001] This invention belongs to the field of low-light image enhancement technology, and relates to a low-light image enhancement method and apparatus based on wavelet frequency domain residual quotient learning. Background Technology

[0002] Low-Light Image Enhancement (LLIE) is a fundamental task in image processing and computer vision, aiming to improve the visibility and perceptual quality of images captured in low-light environments. Low-light images typically suffer from insufficient brightness, color distortion, and noise enhancement, severely impacting human visual perception and the accuracy of subsequent high-level visual tasks, such as object detection, video surveillance, autonomous driving, and medical imaging. Therefore, maintaining structural and color consistency while improving brightness and contrast has always been a crucial research topic in LLIE. Early methods primarily relied on pixel-space grayscale mapping and histogram stretching, such as histogram equalization and gamma correction. These methods are computationally efficient and simple to implement, but they are prone to noise amplification and color distortion under non-uniform lighting conditions. To address these issues, Land et al. proposed the well-known Retinex theory, modeling images as the product of reflection and illumination components. This model has a solid physical foundation and thus serves as a crucial theoretical support for low-light enhancement research.

[0003] With the development of deep learning, end-to-end Convolutional Neural Networks (CNNs) and Transformer structures have been introduced into the LLIE task, greatly improving augmentation performance. For example, Wei et al. proposed RetinexNet, which for the first time achieved end-to-end learning based on Retinex decomposition; Wu et al. further proposed URetinex-Net, which improves illumination estimation accuracy through a learnable iterative unfolding method; and Chen et al. designed EnlightenGAN, which achieves unpaired training through a generative adversarial framework. However, existing models still have the following key limitations:

[0004] (1) Spatial domain convolution coupling problem. Existing networks mainly perform convolution operations in the spatial domain, which makes it difficult to effectively distinguish between high-frequency noise and low-frequency illumination information, resulting in artifacts or over-sharpening of textures in the enhancement results.

[0005] (2) Insufficient utilization of frequency domain features. Most depth enhancement methods only learn brightness mapping at the pixel level, ignoring the patterns of spectral energy distribution. In fact, the main degradation features of low-light images are concentrated in the low-frequency region, while noise is mainly distributed in the high-frequency region. How to utilize frequency domain information for decoupling is the key to improving the interpretability and robustness of the model.

[0006] (3) Zero-reference learning has poor stability. Zero-reference augmentation methods do not require paired data, but their optimization depends on indirect losses (such as exposure or color consistency constraints), the training process is unstable, and the problem of brightness overcompensation is prone to occur. Summary of the Invention

[0007] To address the shortcomings of existing spatial domain low-light enhancement methods in terms of high-low frequency coupling and noise amplification, this invention proposes a low-light image enhancement method and apparatus based on wavelet frequency domain residual quotient learning. This method combines wavelet transform with residual quotient learning, achieving physical interpretability decoupling of illumination and noise in the frequency domain.

[0008] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0009] On the one hand, a low-light image enhancement method based on wavelet frequency domain residual quotient learning is provided, including the following steps:

[0010] Step 1: Acquire and preprocess the low-light image to obtain the low-light input image.

[0011] Step 2: Perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain the low-frequency sub-band and three high-frequency detail sub-bands.

[0012] Step 3: Input the low-frequency sub-band into the low-frequency illumination prediction module to obtain the predicted illumination ratio map. Based on the predicted illumination ratio map and the low-frequency sub-band, obtain the enhanced low-frequency component. The low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency sub-band using a convolutional network structure based on residual learning.

[0013] Step 4: Input the three high-frequency detail subbands and the low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components; the high-frequency suppression module is used to adopt a light-guided high-frequency adaptive suppression mechanism, which dynamically adjusts the high-frequency response through the low-frequency subbands, so that the model can achieve a balance between noise intensity and edge saliency.

[0014] Step 5: Input the enhanced low-frequency component and three enhanced high-frequency components into the image reconstruction module, and reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

[0015] On the other hand, a low-light image enhancement device based on wavelet frequency domain residual quotient learning is also provided, comprising:

[0016] The low-light input image determination unit is used to acquire and preprocess low-light images to obtain low-light input images.

[0017] The wavelet transform unit is used to perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain a low-frequency subband and three high-frequency detail subbands.

[0018] The low-frequency sub-band illumination residual quotient prediction unit is used to input the low-frequency sub-band into the low-frequency illumination prediction module to obtain a predicted illumination ratio map, and to obtain the enhanced low-frequency component based on the predicted illumination ratio map and the low-frequency sub-band; the low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency sub-band using a convolutional network structure based on residual learning.

[0019] The high-frequency suppression unit is used to input three high-frequency detail subbands and low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components. The high-frequency suppression module is used to adopt an illumination-guided high-frequency adaptive suppression mechanism, which dynamically adjusts the high-frequency response through the low-frequency subband, so that the model can achieve a balance between noise intensity and edge saliency.

[0020] The image reconstruction unit is used to input the enhanced low-frequency components and three enhanced high-frequency components into the image reconstruction module, and then reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

[0021] One of the above technical solutions has the following advantages and beneficial effects:

[0022] The aforementioned low-light image enhancement method and apparatus based on wavelet frequency domain residual quotient learning includes: acquiring and preprocessing a low-light image to obtain a low-light input image; performing a two-dimensional discrete wavelet transform on the low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; inputting the low-frequency sub-bands into a low-frequency illumination prediction module to obtain a predicted illumination ratio map; obtaining enhanced low-frequency components based on the predicted illumination ratio map and the low-frequency sub-bands; inputting the three high-frequency detail sub-bands and the low-frequency sub-bands into a high-frequency suppression module to obtain corresponding enhanced high-frequency components; and inputting the enhanced low-frequency components and the three enhanced high-frequency components into an image reconstruction module, performing image reconstruction through inverse discrete wavelet transform to obtain the enhanced low-light image. This method embeds residual quotient learning into the wavelet domain to achieve interpretable enhancement in the frequency dimension. This method inherits the physical consistency of residual quotient learning and introduces the sparse representation capability of the wavelet domain, providing a new perspective for low-light enhancement tasks. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a low-light image enhancement method based on wavelet frequency domain residual quotient learning in one embodiment.

[0025] Figure 2 This is a schematic diagram of the data processing flow in low-light image enhancement based on wavelet frequency domain residual quotient learning in one embodiment.

[0026] Figure 3 This is a schematic diagram of the network structure of a low-frequency illumination prediction module in one embodiment;

[0027] Figure 4 This is a schematic diagram of the network structure of the high-frequency suppression module in one embodiment;

[0028] Figure 5 This is a schematic diagram illustrating the visual contrast of a naturally non-uniformly lit image in the LSRW dataset using different methods in one embodiment. Figure 5 (a) shows the corresponding low-light images input sequentially. Figure 5 (b) to Figure 5 In the middle (i), there are schematic diagrams of visual contrasts of LIME, RetinexNet, KinD, EnlightenGAN, Zero-DCE, RUAS, ResQ-Net, and our method on a natural non-uniformly lit image in the LSRW dataset. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0031] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0032] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0033] In one embodiment, such as Figure 1 , Figure 2 As shown, a low-light image enhancement method based on wavelet frequency domain residual quotient learning is provided, which may include the following processing steps 1 to 5:

[0034] Step 1: Acquire and preprocess the low-light image to obtain the low-light input image.

[0035] Step 2: Perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain the low-frequency sub-band and three high-frequency detail sub-bands.

[0036] Specifically, the changes in illumination in the spatial domain of an image correspond to the low-frequency components in the frequency domain, while details and noise are mainly concentrated in the high-frequency part; frequency domain modeling can naturally achieve the decoupling of illumination and details.

[0037] Wavelet transform, with its excellent time-frequency locality and sparse representation capabilities, is a signal analysis method exhibiting time-frequency locality. It achieves hierarchical image decomposition through multi-scale filtering, striking a balance between spatial and frequency representations. Using Discrete Wavelet Transform (DWT), the image is decomposed into a low-frequency sub-band (Low–Low, LL) and three high-frequency detail sub-bands (Low–High, LH; High–Low, HL; High–High, HH). The low-frequency sub-band primarily contains illumination information, while the high-frequency sub-bands contain texture and noise. This characteristic provides a natural structure for "illumination-noise decoupling" in low-light enhancement.

[0038] This method introduces 2D-DWT to establish a multi-scale feature mapping between space and frequency.

[0039] The 2D-DWT module is used to decompose a low-light input image into a low-frequency subband (LL) and three high-frequency detail subbands (LH, HL, HH), achieving explicit separation of illumination information and texture noise, and effectively decoupling illumination and noise in the frequency domain.

[0040] Step 3: Input the low-frequency sub-band into the low-frequency illumination prediction module to obtain the predicted illumination ratio map; based on the predicted illumination ratio map and the low-frequency sub-band, obtain the enhanced low-frequency component; the low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency sub-band using a convolutional network structure based on residual learning.

[0041] Specifically, by combining discrete wavelet transform with residual quotient learning (RQL), illumination restoration and detail preservation are achieved by performing illumination prediction and noise suppression in the wavelet frequency domain.

[0042] The low-frequency illumination prediction module uses the RQL (Real-Quickness Level) approach to predict illumination ratios in the low-frequency domain. This enables the restoration of illumination and adaptive adjustment of brightness.

[0043] Step 4: Input the three high-frequency detail subbands and the low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components; the high-frequency suppression module is used to adopt a light-guided high-frequency adaptive suppression mechanism, which dynamically adjusts the high-frequency response through the low-frequency subbands, so that the model can achieve a balance between noise intensity and edge saliency.

[0044] Specifically, to address the issue of high-frequency noise amplification, the High-Frequency Suppression Block (HFS module) dynamically suppresses noise and preserves effective texture through a local variance-driven attention mechanism, achieving a balance between structure preservation and detail.

[0045] Step 5: Input the enhanced low-frequency component and three enhanced high-frequency components into the image reconstruction module, and reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

[0046] Specifically, the image reconstruction module constrains the frequency energy distribution of the enhanced image to ensure that the enhancement result remains natural and consistent with the original image in terms of brightness, color, and frequency domain structure.

[0047] The aforementioned low-light image enhancement method and apparatus based on wavelet frequency domain residual quotient learning includes: acquiring and preprocessing a low-light image to obtain a low-light input image; performing a two-dimensional discrete wavelet transform on the low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; inputting the low-frequency sub-bands into a low-frequency illumination prediction module to obtain a predicted illumination ratio map; obtaining enhanced low-frequency components based on the predicted illumination ratio map and the low-frequency sub-bands; inputting the three high-frequency detail sub-bands and the low-frequency sub-bands into a high-frequency suppression module to obtain corresponding enhanced high-frequency components; and inputting the enhanced low-frequency components and the three enhanced high-frequency components into an image reconstruction module, performing image reconstruction through inverse discrete wavelet transform to obtain the enhanced low-light image. This method embeds residual quotient learning into the wavelet domain to achieve interpretable enhancement in the frequency dimension. This method inherits the physical consistency of residual quotient learning and introduces the sparse representation capability of the wavelet domain, providing a new perspective for low-light enhancement tasks.

[0048] In one embodiment, step 2 includes: performing a two-dimensional discrete wavelet transform on the low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; the three high-frequency detail sub-bands include: horizontal edge information, vertical edge information, and diagonal texture and noise components; the specific calculation formula is as follows:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] in, This is the low-frequency subband, which includes the global brightness, illumination structure, and energy distribution of the image; Represents horizontal edge information; Indicates vertical edge information; This indicates the diagonal texture and noise components; Input image for low light conditions. ; and These represent the low-pass and high-pass filter kernels, used to separate low-light input images. and Frequency characteristics of direction; superscript x and y They represent direction and Direction; Symbol This represents a two-dimensional convolution operation.

[0054] Specifically, in low-light images, most of the brightness energy is concentrated in the low-frequency region, while the high-frequency region contains texture, edges, and noise components. Therefore, wavelet decomposition provides a natural frequency decoupling mechanism, allowing the low-frequency subband to... The illumination component in Retinex theory High-frequency subband ( Then the corresponding reflection component The decomposition method allows the model to implement targeted enhancement strategies across different frequency bands: restoring illumination in the low-frequency domain and suppressing noise while preserving details in the high-frequency domain. Compared to traditional spatial domain convolution operations, the feature sparsity of the wavelet domain significantly improves the model's computational efficiency and generalization ability. Furthermore, wavelet transform satisfies the energy conservation property. This property indicates that the energy of the original image remains balanced across frequency bands, a characteristic that ensures the physical interpretability of the enhancement process.

[0055] In one embodiment, the low-frequency illumination prediction module includes: a shallow feature extraction unit, multiple residual blocks, and an output mapping unit; the residual blocks include a first convolutional module and a second convolutional module, and residual learning is achieved through skip connections; the convolutional module includes a first convolutional layer, a batch normalization layer, and a ReLU function; the output mapping unit includes an SE attention module, a second convolutional layer, a ReLU function, a third convolutional layer, and a Sigmoid function; step 3 includes: inputting the low-frequency subband into the shallow feature extraction unit, performing convolution and batch normalization processing, and then applying the ReLU function. Numerical activation is used to obtain shallow features representing global brightness structure information. These shallow features are then progressively enhanced through multiple residual blocks to obtain deep features. The deep features are input into the output mapping unit and processed through the SE attention module, the second convolutional layer, the ReLU function, the third convolutional layer, and the Sigmoid function to obtain the low-frequency residual quotient. The low-light input image is fused with the low-frequency residual quotient via long skip connections to obtain the predicted illumination ratio map. Based on the predicted illumination ratio map and the low-frequency sub-band, the enhanced low-frequency component is obtained. The expression for the enhanced low-frequency component is:

[0056] ;

[0057] in, To enhance the low-frequency components, This represents a plot showing the predicted light ratio. This is a low-frequency sub-band.

[0058] Specifically, according to Retinex theory, the human visual system's perception of an object's brightness is determined by both the object's reflectivity and the lighting conditions, that is:

[0059] ;

[0060] in, Represents the observed image, This is the reflection component (which reflects the properties of the object itself). This refers to the illumination component. The core objective of low-light enhancement is to estimate a reasonable illumination component. And while preserving reflective texture, it restores normal brightness. However, traditional direct prediction in the spatial domain... The method has two problems: (1) the convolution kernel has a limited range of action and it is difficult to capture the global illumination distribution; (2) high-frequency noise is easily mixed with brightness texture in spatial domain features, resulting in artifacts or uneven brightness in the enhancement results. In the existing technology, the illumination estimation problem is redefined as the ratio modeling problem between the input brightness and the target brightness. The relationship between the two is as follows:

[0061] ;

[0062] in, Indicates the brightness ratio. To input a low-light image, This is the enhanced target image. If accurate prediction is possible... The enhancement problem can then be transformed into:

[0063] ;

[0064] This avoids the fuzzy coupling problem inherent in traditional Retinex models during independent decomposition. To further improve training stability and enhance model convergence, RQL rewrites the ratio mapping in residual form:

[0065] ;

[0066] Here, the constant term 1 represents the identity mapping (i.e., no enhancement). This is the residual quotient component of the network prediction.

[0067] In existing technologies, direct regression in the spatial domain In this application, RQL is extended to the wavelet low-frequency domain. In 2D-DWT, the low-frequency subband... Containing the main illumination information of the image, performing quotient learning on this sub-band can effectively avoid high-frequency noise interference. This method defines the low-frequency illumination enhancement process as follows:

[0068] ;

[0069] in, This is a graph showing the predicted illumination ratios generated by the low-frequency illumination prediction module. This is the enhanced low-frequency component.

[0070] The network structure of the low-frequency illumination prediction module is as follows: Figure 3 As shown, the low-frequency illumination prediction module uses a convolutional network structure based on residual learning to model the illumination residual quotient in the low-frequency sub-band. This module consists of a shallow feature extraction unit, multiple residual blocks, and an output mapping unit, improving illumination prediction accuracy while ensuring training stability. First, the low-frequency sub-band is input. Through a layer Convolutional layers, batch normalization layers, and the ReLU function extract basic features to represent global brightness structure information. These features are then progressively enhanced through multiple residual blocks, each consisting of two convolutional layers. Residual learning is achieved through skip connections, effectively mitigating the vanishing gradient problem and improving the modeling ability for complex illumination distributions. After residual feature learning, the module fuses shallow and deep features through long skip connections to preserve the original illumination structure information and avoid over-smoothing. Subsequently, an SE attention module is introduced to adaptively recalibrate channel features, enhancing feature responses related to illumination changes. Finally, the fused features are convolutionally mapped and processed by the Sigmoid function to output a predicted illumination ratio map. (That is, the illumination ratio between the enhanced normal illumination image and the low illumination input image), and constrain its value range to achieve stable and reliable illumination modulation.

[0071] In one embodiment, the high-frequency suppression module includes three parallel branches with identical structures but whose parameters are not shared. Each branch includes a third convolutional module, two first residual modules, a fourth convolutional module, a mapping module composed of lightweight convolutions, and a sigmoid function. The first residual module includes a fourth convolutional layer, a GN layer, and a ReLU function, and residual learning is achieved through skip connections. The third convolutional module includes a fifth convolutional layer, a GN layer, and a ReLU function. The fourth convolutional module includes a sixth convolutional layer, a GN layer, and a ReLU function. The mapping module includes a seventh convolutional layer, a ReLU function, and an eighth convolutional layer. The three high-frequency detail subbands include: horizontal edge information, vertical edge information, and diagonal texture and noise components. Step 4 includes: connecting the low-frequency subbands with the horizontal edge information... The information, vertical edge information, and diagonal texture and noise components are concatenated along the channel dimension and then input into three branches. In each branch, after channel mapping by the third convolution module, they are modeled by two first residual modules. The modeling results are processed by the fourth convolution module and the mapping module, and then activated by the Sigmoid function to generate the corresponding illumination modulation gate weight map. The horizontal edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced horizontal edge information. The vertical edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced vertical edge information. The diagonal texture and noise components are multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced diagonal texture and noise components.

[0072] Specifically, after wavelet decomposition, three high-frequency detail sub-bands ( The high-frequency features of an image contain both texture and noise components. Directly enhancing these features would amplify the noise signal, leading to artifacts, edge breaks, and local oversharpening. Therefore, our method designs a high-frequency suppression module to adaptively suppress noise energy while preserving structural details. The high-frequency components in the wavelet domain exhibit significant differences in their response to brightness and texture gradients in different regions. Their statistical characteristics can be approximated by a local variance model; low-variance regions often correspond to noise-dominated smooth areas, while high-variance regions represent significant edges or detailed structures. Based on this characteristic, we propose an illumination-guided high-frequency adaptive suppression mechanism that utilizes low-frequency illumination information... The high-frequency response is dynamically adjusted to achieve a balance between noise intensity and edge saliency in the model. The specific process is as follows: First, for each high-frequency sub-band... Combine it with low-frequency light-irradiated subband Perform channel splicing:

[0073] ;

[0074] The spliced ​​features The intermediate features are obtained after processing by the third convolutional module, two first residual modules, and the fourth convolutional module. Then the intermediate features Joint contextual information is extracted through multi-layer convolution and nonlinear mapping to generate a gating graph for illumination modulation. :

[0075] ;

[0076] in, The sigmoid function is used to restrict the output to a range. Subsequently, adaptive weighted suppression is implemented using a gating mechanism:

[0077] ;

[0078] Among them, the symbol " " indicates element-wise multiplication. When When the frequency is low, the module automatically attenuates high-frequency energy to suppress interference in dark or noise-dominated areas; while when... When the response is large, it is maintained or amplified at the edges and in high-reflection areas, thereby achieving the goal of noise suppression and structural preservation in parallel.

[0079] The network structure of the high-frequency suppression module is as follows: Figure 4 As shown, the high-frequency suppression module consists of three parallel branches with identical structures but whose parameters are not shared, each corresponding to a high-frequency sub-band. , , For each high-frequency subband First, it is compared with the low-frequency subband. The concatenated features are used as input along the channel dimension. After channel mapping by convolutional layers, the concatenated features are input into a feature extraction network composed of multiple residual blocks for modeling. Subsequently, the features are further integrated through convolutional layers and input to the corresponding output branches. Each branch consists of a lightweight convolutional mapping module, and a gated weight map is generated using the Sigmoid function. Ultimately, each high-frequency subband is modulated using an element-wise weighted method. The three branches maintain a consistent structural form, but their parameters are learned independently, thereby achieving differentiated modeling of high-frequency characteristics in different directions.

[0080] In one embodiment, the low-light image enhancement model based on wavelet frequency domain residual quotient learning includes a 2D-DWT module, a low-frequency illumination prediction module, a high-frequency suppression module, and an image reconstruction module; the expression for the overall loss function used in the training process of the low-light image enhancement model based on wavelet frequency domain residual quotient learning is as follows:

[0081] ;

[0082] in, For the overall loss function, , , and These represent reconstruction consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss, respectively. The weighting coefficients for reconstructing consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss are used to balance the optimization relationship between different constraints.

[0083] Specifically, under unsupervised pairing, low-light image enhancement tasks lack high-quality reference images as direct optimization targets, easily leading to problems such as brightness overcompensation, structural distortion, and color drift during network training. To ensure the convergence stability of the model and the naturalness of the enhancement results in zero-reference scenarios, this method, based on the residual quotient learning framework and wavelet frequency domain modeling, comprehensively draws on the constraint ideas from RetinexNet, ResQ-Net, Zero-DCE, and existing frequency domain enhancement methods. It designs a multi-constraint joint optimization no-reference loss function system, constraining the network from multiple dimensions such as physical consistency, exposure control, frequency domain structure preservation, and color constancy. The overall loss function is defined as shown in the expression above.

[0084] Taking into account the physical meaning and numerical scale differences of each loss term, this embodiment determines the loss weight parameters based on experience as follows: , , , The network was trained using the Adam optimizer, with an initial learning rate set to... Furthermore, a cosine annealing strategy is employed to gradually decay the model, thereby improving its convergence stability and generalization ability.

[0085] In one embodiment, the reconstruction consistency loss is used to constrain the multiplicative consistency relationship between the enhancement result and the quotient mapping; the expression for the reconstruction consistency loss is:

[0086] ;

[0087] in, For the enhanced low-light image, This indicates that the input image is a low-light image. This represents element-wise multiplication. To predict the light ratio map, This represents the L1 norm.

[0088] Specifically, based on Retinex theory and the residual quotient learning model, the enhanced low-light image... Plot of ratio of predicted illumination Physical imaging relationship should be satisfied ,in This represents the input low-light image. Referring to the reconstruction constraint concept based on the imaging model in RetinexNet and ResQ-Net, this method introduces a reconstruction consistency loss to constrain the multiplicative consistency relationship between the augmentation result and the quotient mapping. Its definition is shown in the expression for the reconstruction consistency loss above. This loss term effectively suppresses unreasonable brightness amplification and structural distortion by constraining the joint reconstruction error between the augmented image and the predicted illumination ratio map, thereby enhancing the physical interpretability and training stability of the model.

[0089] In one embodiment, the exposure control loss is used to constrain the enhancement magnitude of low-brightness regions; the exposure control loss is:

[0090] ;

[0091] ;

[0092] in, Indicates dark area mask, Indicates pixel position, The threshold for determining dark areas. Indicates the desired brightness level. For the enhanced low-light image (enhanced low-light image) These are the numerical values ​​of the image during the computation process; an image is actually a three-channel matrix. This represents element-wise multiplication. pixel position Input low-light image, This represents the L1 norm.

[0093] Specifically, in zero-reference enhancement, the network is prone to over-enhancing in dark areas, leading to problems such as local overexposure, loss of detail, and unnatural visual appearance. To avoid these issues, this method borrows the exposure adjustment strategies from Zero-DCE and RUAS, introducing dark area exposure control loss to constrain the enhancement amplitude in low-brightness areas. First, a dark area mask is constructed based on the brightness information of the low-light input image. Based on this, the exposure control loss is defined as shown in the expression for the reconstruction consistency loss above. This loss term, by imposing targeted constraints on dark areas, guides the network to adaptively adjust the enhancement intensity of low-brightness areas, effectively avoiding local overexposure while improving visibility.

[0094] In one embodiment, frequency domain consistency loss is used to constrain the high-frequency energy changes before and after enhancement; the expression for frequency domain consistency loss is:

[0095] ;

[0096] ;

[0097] in, , These are the high-frequency sub-bands before and after enhancement, This indicates the total number of pixels.

[0098] Specifically, this method models images in the frequency domain based on wavelet transform, where the high-frequency detail sub-bands mainly carry the image's texture structure and noise information. Without effective constraints, the enhancement process can easily lead to abnormal amplification or excessive suppression of high-frequency energy, resulting in over-sharpening or over-smoothing. Referring to the energy constraint concept in existing techniques for frequency domain enhancement, this method introduces frequency domain consistency loss in the wavelet domain to constrain the changes in high-frequency energy before and after enhancement. Let the high-frequency sub-bands before and after enhancement be... and Its corresponding frequency domain energy is defined as: The frequency domain consistency loss is defined as shown in the formula for the frequency domain consistency loss. This constraint enables the network to maintain the relative stability of high-frequency energy distribution during the enhancement process, effectively suppressing noise amplification and texture degradation, and improving structural fidelity.

[0099] In one embodiment, color consistency loss is used to constrain the RGB channels; the expression for color consistency loss is:

[0100] ;

[0101] in, These represent the average values ​​of the three color channels of the enhanced image. This is the global average. .

[0102] Specifically, in low-light image enhancement, unbalanced amplification between different color channels can easily cause color cast, affecting visual realism. To maintain color constancy in the enhanced image, this method, based on the color constraint concept in Zero-DCE, introduces color consistency loss to constrain the RGB channels. Let the average values ​​of the three color channels of the enhanced image be... Its global average value is defined as The color consistency loss is defined as shown in the expression for color consistency loss. This loss term effectively mitigates color imbalance that may occur during the enhancement process and improves the color naturalness of the enhancement result by constraining the statistical consistency relationship between different color channels.

[0103] In a validation embodiment, to fully verify the effectiveness of this method in low-light image enhancement tasks, systematic experiments were conducted on the LOL, SICE, and LSRW public datasets. The model performance was evaluated from both quantitative and qualitative perspectives. This method is implemented based on the PyTorch deep learning framework, and all experiments were performed on a computing platform equipped with an NVIDIA GPU 5060Ti. During model training, all low-light input images were uniformly scaled to a fixed resolution and subjected to data augmentation operations such as random cropping and flipping to improve the model's generalization ability. The network training used the Adam optimizer, with the following parameter settings: initial learning rate... , , The weight decay coefficient is During training, a cosine annealing learning rate decay strategy is employed, gradually decreasing the learning rate with each training epoch to improve convergence accuracy in later stages. The batch size is set to 8, and the model is trained for 400 epochs. Training is terminated early when the validation set loss no longer decreases significantly to avoid overfitting.

[0104] (1) Evaluation indicators

[0105] To objectively evaluate the performance of different low-light enhancement methods, this method comprehensively analyzes the results from two levels: reference-based evaluation metrics and reference-free perceptual evaluation metrics. For datasets with paired reference images, this embodiment uses Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as the main evaluation metrics. PSNR measures the pixel reconstruction error between the enhanced image and the reference image; SSIM characterizes image similarity from three dimensions: brightness, contrast, and structure, which better aligns with the structural sensitivity of the visual system; LPIPS utilizes deep convolutional feature space modeling, which more accurately reflects human subjective visual perception than traditional metrics. Furthermore, to assess the visual naturalness of the images, this method introduces the Natural Image Quality Evaluator (NIQE). NIQE builds a model based on the statistical characteristics of natural scenes, allowing evaluation of the naturalness and artifact levels of the enhancement results without the need for a reference image. The above metrics comprehensively reflect the overall performance of different enhancement algorithms from multiple aspects, including pixel accuracy, structure preservation, and perception quality.

[0106] (2) Comparative analysis of experimental results

[0107] To comprehensively evaluate the performance of the proposed method, extensive comparative experiments were conducted on three representative benchmark datasets: the LOL dataset (including v1 and v2), the LSRW dataset (large-scale real-world scenes), and the SICE dataset (multi-exposure sequences). The proposed method was compared with seven mainstream methods, covering traditional methods like LIME, supervised learning methods like RetinexNet and KinD, and unsupervised / zero-reference methods like EnlightenGAN, Zero-DCE, RUAS, and ResQ-Net. Evaluation metrics included the full-reference metrics PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity), and LPIPS (Perceptual Patch Similarity), and the no-reference metric NIQE (Natural Image Quality Evaluation). Higher PSNR and SSIM scores were preferred, while lower LPIPS and NIQE scores were preferred.

[0108] (a) Quantitative comparison results

[0109] Evaluation results of the LOL dataset (v1+v2): The LOL dataset contains pairs of low-light / normal-light images from real-world scenes. Considering the scene differences between LOL datasets v1 and v2, the two datasets were merged (a total of 115 test images) for unified testing, and the results are shown in Table 1.

[0110] Table 1: Quantitative comparison results on the LOL dataset (v1+v2)

[0111]

[0112] As can be seen from the table, KinD, as a representative of supervised learning methods, achieves the best performance in PSNR and SSIM due to the strong pixel-level constraints provided by paired data. However, under the more challenging unsupervised / zero-reference setting, our method demonstrates significant advantages. Specifically, our method achieves a PSNR of 19.98 dB, steadily surpassing the state-of-the-art method ResQ-Net (19.71 dB). This indicates that even in the absence of reference image supervision, the residual quotient learning strategy proposed in our method can accurately model the illumination distribution. A more crucial breakthrough lies in the SSIM metric, where our method achieves a high score of 0.785, an improvement of approximately 6.9% compared to ResQ-Net (0.734), and extremely close to the supervised method KinD. This significant improvement is directly attributed to the wavelet transform module introduced in our method: by decoupling high-frequency noise from texture details in the frequency domain, our method effectively overcomes the defect of pure spatial domain methods (such as ResQ-Net) that often sacrifice texture details during smoothing denoising, achieving the best balance between "denoising" and "structure preservation". Furthermore, the optimal NIQE score further confirms that the enhanced image is visually clearer and more natural, with fewer artifacts.

[0113] Generalization test: To verify the robustness of the model in real-world complex scenarios, tests were conducted on the LSRW and SICE datasets, and the results are shown in Table 2.

[0114] Table 2 Quantitative comparison results between LSRW and SICE datasets

[0115]

[0116] The results in Table 2 validate the generalization ability of our proposed method in real-world complex scenes. On the highly challenging LSRW dataset, most methods struggle to maintain image structural integrity due to the test images originating from unpaired, real-world low-light environments. While ResQ-Net boasts a high PSNR, its SSIM score is only 0.471, even lower than the earlier EnlightenGAN. In contrast, our proposed method achieves the highest SSIM score of 0.542, surpassing the supervised learning method KinD. This demonstrates that in real-world scenes lacking paired reference image constraints, the framework proposed in this application exhibits stronger environmental adaptability thanks to the robust extraction capability of high-frequency features using wavelet transform. On the SICE multi-exposure sequence dataset, although ResQ-Net has a slight advantage in PSNR, our proposed method still leads in SSIM, proving that it can recover details more finely when handling high dynamic range scenes.

[0117] (b) Visual contrast results analysis

[0118] To further analyze the visual performance of different low-light enhancement methods in real-world scenes, this embodiment selects typical samples from the LSRW dataset for subjective comparison. The results are as follows: Figure 5 As shown, Figure 5 (a) shows the corresponding low-light images input sequentially. Figure 5 Figures (b) to (i) show a subjective comparison of typical samples from the LSRW dataset using LIME, RetinexNet, KinD, EnlightenGAN, Zero-DCE, RUAS, ResQ-Net, and our proposed method. The red box area is a magnified view of local details, used to compare the differences between the different methods in terms of brightness restoration, structure preservation, and noise suppression.

[0119] From a visual perspective, while LIME improves overall brightness, it suffers from amplified noise in dark areas and blurred details. RetinexNet achieves a relatively uniform brightness distribution, but local structures remain unstable. KinD performs well in terms of brightness and contrast, but exhibits some over-enhancement in highlight areas. EnlightenGAN and Zero-DCE improve visibility, but still suffer from structural blurring or color shift in detail areas, while RUAS's enhancement is relatively limited, resulting in an overall darker image. ResQ-Net has advantages in brightness stability, but some smoothing of local details remains. In contrast, our proposed method better preserves structural details while improving overall brightness and effectively suppresses noise amplification in dark areas. The enhancement results show clearer and more natural edge and texture distribution in local areas, consistent with the quantitative improvement trends of indicators such as SSIM and NIQE in Table 1.

[0120] It should be understood that, although the above Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0121] In one embodiment, a low-light image enhancement device based on wavelet frequency domain residual quotient learning is also provided, comprising:

[0122] The low-light input image determination unit is used to acquire and preprocess low-light images to obtain low-light input images.

[0123] The wavelet transform unit is used to perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain a low-frequency subband and three high-frequency detail subbands.

[0124] The low-frequency sub-band illumination residual quotient prediction unit is used to input the low-frequency sub-band into the low-frequency illumination prediction module to obtain a predicted illumination ratio map, and to obtain the enhanced low-frequency component based on the predicted illumination ratio map and the low-frequency sub-band; the low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency sub-band using a convolutional network structure based on residual learning.

[0125] The high-frequency suppression unit is used to input three high-frequency detail subbands and low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components. The high-frequency suppression module is used to adopt an illumination-guided high-frequency adaptive suppression mechanism, which dynamically adjusts the high-frequency response through the low-frequency subband, so that the model can achieve a balance between noise intensity and edge saliency.

[0126] The image reconstruction unit is used to input the enhanced low-frequency components and three enhanced high-frequency components into the image reconstruction module, and then reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

[0127] In one embodiment, the wavelet transform unit is further used to perform a two-dimensional discrete wavelet transform on the low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; the three high-frequency detail sub-bands include: horizontal edge information, vertical edge information, and diagonal texture and noise components; the specific calculation formula is as follows:

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] in, This is the low-frequency subband, which includes the global brightness, illumination structure, and energy distribution of the image; Represents horizontal edge information; Indicates vertical edge information; This indicates the diagonal texture and noise components; Input image for low light conditions. ; and These represent the low-pass and high-pass filter kernels, used to separate low-light input images. and Frequency characteristics of direction; superscript x and y They represent direction and Direction; Symbol This represents a two-dimensional convolution operation.

[0133] In one embodiment, the low-frequency illumination prediction module includes: a shallow feature extraction unit, multiple residual blocks, and an output mapping unit; the residual blocks include a first convolutional module and a second convolutional module, and residual learning is achieved through skip connections; the convolutional modules include a first convolutional layer, a batch normalization layer, and a ReLU function; the output mapping unit includes an SE attention module, a second convolutional layer, a ReLU function, a third convolutional layer, and a Sigmoid function; the low-frequency subband illumination residual quotient prediction unit is further used to input the low-frequency subband into the shallow feature extraction unit, and after convolution and batch normalization... After standardization, ReLU activation is applied to obtain shallow features representing global brightness structure information. These shallow features are then progressively enhanced through multiple residual blocks to obtain deep features. These deep features are input into the output mapping unit and processed through the SE attention module, the second convolutional layer, the ReLU function, the third convolutional layer, and the Sigmoid function to obtain the low-frequency residual quotient. The low-illuminance input image is fused with the low-frequency residual quotient via long skip connections to obtain the predicted illumination ratio map. Based on the predicted illumination ratio map and the low-frequency sub-band, the enhanced low-frequency components are obtained as follows:

[0134] ;

[0135] in, To enhance the low-frequency components, This represents a plot showing the predicted light ratio. This is a low-frequency sub-band.

[0136] In one embodiment, the high-frequency suppression module includes three parallel branches with identical structures but whose parameters are not shared. Each branch includes a third convolutional module, two first residual modules, a fourth convolutional module, a mapping module composed of lightweight convolutions, and a sigmoid function. The first residual module includes a fourth convolutional layer, a GN layer, and a ReLU function, and residual learning is achieved through skip connections. The third convolutional module includes a fifth convolutional layer, a GN layer, and a ReLU function. The fourth convolutional module includes a sixth convolutional layer, a GN layer, and a ReLU function. The mapping module includes a seventh convolutional layer, a ReLU function, and an eighth convolutional layer. The three high-frequency detail subbands include: horizontal edge information, vertical edge information, and diagonal texture and noise components. The high-frequency suppression unit is also used to separate the low-frequency subbands from the horizontal edge information and the vertical edge information. Edge information, vertical edge information, and diagonal texture and noise components are concatenated along the channel dimension and then input into three branches. In each branch, after channel mapping by the third convolution module, they are modeled by two first residual modules. The modeling results are processed by the fourth convolution module and the mapping module, and then activated by the Sigmoid function to generate the corresponding illumination modulation gate weight map. The horizontal edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced horizontal edge information. The vertical edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced vertical edge information. The diagonal texture and noise components are multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced diagonal texture and noise components.

[0137] In one embodiment, the low-light image enhancement model based on wavelet frequency domain residual quotient learning includes a 2D-DWT module, a low-frequency illumination prediction module, a high-frequency suppression module, and an image reconstruction module; the overall loss function used in the training process of the low-light image enhancement model based on wavelet frequency domain residual quotient learning is:

[0138] ;

[0139] in, For the overall loss function, , , and These represent reconstruction consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss, respectively. The weighting coefficients for reconstructing consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss are used to balance the optimization relationship between different constraints.

[0140] In one embodiment, the reconstruction consistency loss is used to constrain the multiplicative consistency relationship between the enhancement result and the quotient mapping; the reconstruction consistency loss is:

[0141] ;

[0142] in, For the enhanced low-light image, This indicates that the input image is a low-light image. This represents element-wise multiplication. To predict the light ratio map, This represents the L1 norm.

[0143] In one embodiment, the exposure control loss is used to constrain the enhancement magnitude of low-brightness regions; the exposure control loss is:

[0144] ;

[0145] ;

[0146] in, Indicates dark area mask, Indicates pixel position, The threshold for determining dark areas. Indicates the desired brightness level. For the enhanced low-light image, This represents element-wise multiplication. pixel position Input low-light image, This represents the L1 norm.

[0147] In one embodiment, the frequency domain consistency loss is used to constrain the high-frequency energy changes before and after enhancement; the frequency domain consistency loss is:

[0148] ;

[0149] ;

[0150] in, , These are the high-frequency sub-bands before and after enhancement, This indicates the total number of pixels.

[0151] In one embodiment, color consistency loss is used to constrain the RGB channels; the color consistency loss is:

[0152] ;

[0153] in, These represent the average values ​​of the three color channels of the enhanced image. This is the global average. .

[0154] It is understood that for a detailed explanation of the low-light image enhancement device based on wavelet frequency domain residual quotient learning, please refer to the corresponding explanations of the various embodiments of the low-light image enhancement method based on wavelet frequency domain residual quotient learning above, and will not be repeated here. Each module in the aforementioned low-light image enhancement device based on wavelet frequency domain residual quotient learning can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer devices already existing in the art.

[0155] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0156] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A low-light image enhancement method based on wavelet frequency domain residual quotient learning, characterized in that, Including the following steps: Step 1: Acquire and preprocess the low-light image to obtain the low-light input image; Step 2: Perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; Step 3: Input the low-frequency sub-band into the low-frequency illumination prediction module to obtain the predicted illumination ratio map. Based on the predicted illumination ratio map and the low-frequency sub-band, obtain the enhanced low-frequency component. The low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency subband using a convolutional network structure based on residual learning. The low-frequency illumination prediction module includes: a shallow feature extraction unit, multiple residual blocks, and an output mapping unit; the residual blocks include a first convolutional module and a second convolutional module, and residual learning is achieved through skip connections; the convolutional module includes a first convolutional layer, a batch normalization layer, and a ReLU function; the output mapping unit includes an SE attention module, a second convolutional layer, a ReLU function, a third convolutional layer, and a Sigmoid function; Step 4: Input the three high-frequency detail subbands and the low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components. The high-frequency suppression module uses an illumination-guided high-frequency adaptive suppression mechanism to dynamically adjust the high-frequency response through the low-frequency subbands, enabling the model to achieve a balance between noise intensity and edge saliency. The high-frequency suppression module includes three parallel branches with identical structures but non-shared parameters. Each branch includes a third convolutional module, two first residual modules, a fourth convolutional module, a mapping module composed of lightweight convolutions, and a Sigmoid function. The first residual module includes a fourth convolutional layer, a GN layer, and a ReLU function, and achieves residual learning through skip connections. The third convolutional module includes a fifth convolutional layer, a GN layer, and a ReLU function. The fourth convolutional module includes a sixth convolutional layer, a GN layer, and a ReLU function. The mapping module includes a seventh convolutional layer, a ReLU function, and an eighth convolutional layer. Step 5: Input the enhanced low-frequency component and three enhanced high-frequency components into the image reconstruction module, and reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

2. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 1, characterized in that, Step 2 includes: performing a two-dimensional discrete wavelet transform on the low-light input image using a 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands; the three high-frequency detail sub-bands include: horizontal edge information, vertical edge information, and diagonal texture and noise components; the specific calculation formula is as follows: in, This is the low-frequency subband, which includes the global brightness, illumination structure, and energy distribution of the image; Represents horizontal edge information; Indicates vertical edge information; This indicates the diagonal texture and noise components; Input image for low light conditions. ; and These represent the low-pass and high-pass filter kernels, used to separate low-light input images. and Frequency characteristics of direction; superscript x and y They represent direction and Direction; Symbol This represents a two-dimensional convolution operation.

3. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 1, characterized in that, Step 3 includes: The low-frequency subband is input into the shallow feature extraction unit, and after convolution and batch normalization, it is activated by the ReLU function to obtain shallow features that characterize global brightness structure information. The shallow features are enhanced step by step through multiple residual blocks to obtain deep features; The deep features are input into the output mapping unit, and processed by the SE attention module, the second convolutional layer, the ReLU function, the third convolutional layer, and the Sigmoid function to obtain the low-frequency residual quotient. The low-light input image is fused with the low-frequency residual quotient through long-skip connections to obtain a predicted illumination ratio map. Based on the predicted illumination ratio map and the low-frequency sub-band, the enhanced low-frequency component is obtained as follows: in, To enhance the low-frequency components, This represents a plot showing the predicted light ratio. This is a low-frequency sub-band.

4. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 1, characterized in that, The three high-frequency detail subbands include: horizontal edge information, vertical edge information, and diagonal texture and noise components; Step 4 includes: The low-frequency subband is concatenated with the horizontal edge information, the vertical edge information, and the diagonal texture and noise components in the channel dimension and then input into three branches respectively. In each branch, the channel is mapped by the third convolution module and then modeled by two first residual modules. The modeling results are processed by the fourth convolution module and the mapping module, and then activated by the Sigmoid function to generate the corresponding illumination modulation gating weight map. The horizontal edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced horizontal edge information; The vertical edge information is multiplied element-wise with the corresponding illumination modulation gate weight map to obtain the enhanced vertical edge information. The diagonal texture and noise components are multiplied element-wise with the corresponding illumination modulation gating weight map to obtain the enhanced diagonal texture and noise components.

5. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 1, characterized in that, The low-light image enhancement model based on wavelet frequency domain residual quotient learning includes a 2D-DWT module, a low-frequency illumination prediction module, a high-frequency suppression module, and an image reconstruction module. The overall loss function used in the training process of the low-light image enhancement model based on wavelet frequency domain residual quotient learning is: in, For the overall loss function, , , and These represent reconstruction consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss, respectively. The weighting coefficients for reconstructing consistency loss, exposure control loss, frequency domain consistency loss, and color consistency loss are used to balance the optimization relationship between different constraints.

6. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 5, characterized in that, The reconstruction consistency loss is used to constrain the multiplicative consistency relationship between the enhancement result and the quotient mapping; the reconstruction consistency loss is: in, For the enhanced low-light image, This indicates that the input image is a low-light image. This represents element-wise multiplication. To predict the illumination ratio map, This represents the L1 norm.

7. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 5, characterized in that, The exposure control loss is used to constrain the enhancement magnitude of low-brightness areas; the exposure control loss is: in, Indicates dark area mask, Indicates pixel position, The threshold for determining dark areas. Indicates the desired brightness level. For the enhanced low-light image, This represents element-wise multiplication. pixel position Input low-light image, This represents the L1 norm.

8. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 5, characterized in that, The frequency domain consistency loss is used to constrain the high-frequency energy changes before and after enhancement; the frequency domain consistency loss is: in, , These are the high-frequency sub-bands before and after enhancement, This indicates the total number of pixels.

9. The low-light image enhancement method based on wavelet frequency domain residual quotient learning according to claim 5, characterized in that, The color consistency loss is used to constrain the RGB channels; the color consistency loss is: in, These represent the average values ​​of the three color channels of the enhanced image. This is the global average. .

10. A low-light image enhancement device based on wavelet frequency domain residual quotient learning, characterized in that, include: The low-light input image determination unit is used to acquire and preprocess a low-light image to obtain a low-light input image; The wavelet transform unit is used to perform two-dimensional discrete wavelet transform on the low-light input image using the 2D-DWT module to obtain a low-frequency sub-band and three high-frequency detail sub-bands. The low-frequency sub-band illumination residual quotient prediction unit is used to input the low-frequency sub-band into the low-frequency illumination prediction module to obtain a predicted illumination ratio map, and to obtain the enhanced low-frequency component based on the predicted illumination ratio map and the low-frequency sub-band. The low-frequency illumination prediction module is used to predict the illumination residual quotient in the low-frequency subband using a convolutional network structure based on residual learning. The low-frequency illumination prediction module includes: a shallow feature extraction unit, multiple residual blocks, and an output mapping unit; the residual blocks include a first convolutional module and a second convolutional module, and residual learning is achieved through skip connections; the convolutional module includes a first convolutional layer, a batch normalization layer, and a ReLU function; the output mapping unit includes an SE attention module, a second convolutional layer, a ReLU function, a third convolutional layer, and a Sigmoid function; A high-frequency suppression unit is used to input three high-frequency detail subbands and the low-frequency subbands into the high-frequency suppression module to obtain the corresponding enhanced high-frequency components. The high-frequency suppression module uses an illumination-guided high-frequency adaptive suppression mechanism to dynamically adjust the high-frequency response through the low-frequency subbands, enabling the model to achieve a balance between noise intensity and edge saliency. The high-frequency suppression module includes three parallel branches with identical structures but non-shared parameters. Each branch includes a third convolutional module, two first residual modules, a fourth convolutional module, a mapping module composed of lightweight convolutions, and a sigmoid function. The first residual module includes a fourth convolutional layer, a GN layer, and a ReLU function, and achieves residual learning through skip connections. The third convolutional module includes a fifth convolutional layer, a GN layer, and a ReLU function. The fourth convolutional module includes a sixth convolutional layer, a GN layer, and a ReLU function. The mapping module includes a seventh convolutional layer, a ReLU function, and an eighth convolutional layer. The image reconstruction unit is used to input the enhanced low-frequency components and three enhanced high-frequency components into the image reconstruction module, and then reconstruct the image through discrete wavelet inverse transform to obtain the enhanced low-light image.

Images