A low-light image enhancement method based on wavelet space dynamic enhancement network

Abstract

The application discloses a low-light image enhancement method based on a wavelet space dynamic enhancement network. In view of the fact that the existing low-light enhancement method based on a wavelet is prone to causing problems such as unstable low-frequency brightness baseline, unnatural brightness transition and low-frequency artifacts and the like due to the overall and coarse-grained enhancement of the low-frequency subband before fine high-frequency recovery, a low-frequency dynamic separation enhancement module is arranged at a low-frequency processing link in a wavelet domain, low-frequency features are separated into a smooth low-frequency response and a low-frequency residual response, and controlled enhancement is realized through selective modulation and gated residual updating; meanwhile, a scale perception spatial convolution strategy is arranged at a spatial domain compensation link after inverse wavelet reconstruction, fixed overlapping grouped convolution or ordinary spatial convolution is selected according to feature scales, and the strategy is used to reduce redundant parameters and calculation overhead. The application can improve the brightness recovery stability and visual naturalness of low-light images and reduce the model complexity.

Description

Technical Field

[0001] This invention relates to low-light image enhancement, image restoration and image quality improvement in computer vision, and particularly to a low-light image enhancement method based on wavelet domain low-frequency dynamic separation enhancement and fixed overlap grouped convolution lightweighting. Background Technology

[0002] Low-light image enhancement is a crucial image restoration task in computer vision and image processing, aiming to improve the visibility and perceptual quality of images in low-light environments. Images acquired in real-world low-light scenes often suffer from degradation issues such as insufficient brightness, low contrast, color shift, amplified noise, and weakened local structures. These problems not only affect human visual perception but also reduce the reliability of subsequent visual tasks such as object detection, image recognition, semantic segmentation, intelligent surveillance, and autonomous driving. Therefore, how to restore natural brightness, maintain color stability, and reconstruct structural details in complex low-light scenes is a critical problem that needs to be solved in the field of low-light image enhancement.

[0003] Existing low-light image enhancement methods mainly include traditional image enhancement methods and deep learning-based methods. Traditional methods typically include histogram equalization, gamma correction, and enhancement methods based on Retinex theory, which mainly achieve image brightness enhancement through pixel distribution adjustment, brightness mapping, or illumination estimation. These methods have relatively simple structures, but usually rely on manually designed priors or fixed enhancement rules, and are prone to problems such as over-enhancement, color distortion, noise amplification, and loss of local details in complex real-world low-light scenes. With the development of deep learning, methods such as convolutional neural networks, Transformers, and state-space models have been gradually applied to low-light image enhancement tasks. These methods can learn the mapping relationship between low-light images and normal-light images in a data-driven manner, and have made significant progress in brightness restoration and detail enhancement. However, under complex low-light degradation conditions, natural brightness restoration, color stability preservation, and structural detail reconstruction remain challenging.

[0004] In recent years, frequency domain modeling methods have provided new approaches for low-light image enhancement. These methods can decompose image features into different frequency components, thus facilitating the relatively decoupled modeling of brightness information and detail components. Among them, the Fourier transform has global frequency modeling capabilities, making it suitable for handling overall brightness-related information; however, it has limitations in processing local structural details due to the lack of explicit spatial localization capabilities. In contrast, the wavelet transform can decompose image features into low-frequency sub-bands and multiple directional high-frequency sub-bands, preserving spatial relationships while performing frequency decomposition. Therefore, it is more suitable for separately modeling low-frequency brightness baselines and high-frequency edge, texture, and directional detail information.

[0005] However, existing wavelet domain low-light image enhancement methods still have shortcomings. Some methods, such as Wave-Mamba (W. Zou, H. Gao, W. Yang, T. Liu, Wave-Mamba: Wavelet State Space Model for Ultra-High-Definition Low-Light Image Enhancement, Proc. ACM Int. Conf. Multimedia, pp. 1534–1543, 2024), typically prioritize the recovery of low-frequency subbands and use the enhanced low-frequency features to guide subsequent high-frequency detail reconstruction, but detail recovery is insufficient. Recent methods, such as CWNet (T. Zhang, P. Liu, et al., CWNet: Causal Wavelet Network for Low-Light Image Enhancement, Proc. IEEE / CVFInt. Conf. Comput. Vis., pp. 8789–8799, 2025), further emphasize the importance of directional edges, textures, and structural details in high-frequency subbands for clear and natural enhancement results. However, in processing workflows focused on refined high-frequency restoration, low-frequency subbands are often still treated as a whole and enhanced with relatively coarse-grained precision. In reality, the low-frequency subband is not simply a representation of illumination; it simultaneously contains brightness distribution, color information, structural cues, and low-frequency degradation components, exhibiting significant coupling characteristics. If the low-frequency subband lacks dynamic separation and selective enhancement mechanisms, it can easily introduce an unstable low-frequency brightness baseline, causing subsequent high-frequency detail restoration and spatial structure reconstruction to be built on an unreliable low-frequency foundation. This results in unnatural brightness transitions, color shifts, and low-frequency artifacts in the enhancement results.

[0006] Furthermore, after wavelet domain enhancement, the spatial domain features obtained from inverse wavelet reconstruction still require further compensation and refinement to alleviate potential structural discontinuities and insufficient local texture recovery. Existing methods typically employ standard spatial convolution for local context modeling during the spatial domain reconstruction compensation stage. However, using conventional spatial convolution across all scales introduces parameter redundancy and computational overhead, especially at high resolutions. While ordinary grouped convolution can reduce the number of parameters and computational cost, its hard partitioning of channels into non-overlapping groups can lead to insufficient inter-group information exchange, weakening the spatial feature representation capability. Therefore, a low-light image enhancement method is urgently needed that can rationally allocate spatial modeling capabilities according to the feature recovery requirements at different scales, reducing redundant computation while maintaining spatial reconstruction capabilities.

[0007] Based on the above problems, there is an urgent need to propose a new low-light image enhancement method that can address the coupling between brightness distribution, main structural cues, and low-frequency degradation components in wavelet domain low-frequency subband processing. This method should dynamically separate and selectively enhance the smooth brightness components and structural residual information in the low-frequency subband to establish a more stable low-frequency brightness baseline and improve the stability of overall brightness restoration and main structure preservation. Simultaneously, in the spatial domain compensation stage after inverse wavelet reconstruction, spatial modeling capabilities need to be rationally allocated according to the restoration requirements of different scale features. This avoids redundant parameters and computational overhead caused by using conventional spatial convolution at all scales, thereby reducing model complexity while maintaining local spatial context modeling capabilities. Summary of the Invention

[0008] This invention proposes a low-light image enhancement method based on a wavelet spatial dynamic enhancement network, addressing the following technical problems: Existing wavelet domain low-light image enhancement methods often perform overall, coarse-grained enhancement of low-frequency subbands before fine-grained high-frequency restoration, lacking a dynamic separation and selective enhancement mechanism for smooth low-frequency components and low-frequency residual information in low-frequency features. This easily leads to instability in the low-frequency brightness baseline, further causing problems such as unnatural brightness transitions, color shifts, and low-frequency artifacts. At the same time, in the spatial domain compensation process after inverse wavelet reconstruction, if conventional spatial convolution is used at all scales, redundant parameters and computational overhead are easily introduced, making it difficult to balance local spatial context modeling capabilities and model complexity control.

[0009] This invention provides a low-light image enhancement method based on a wavelet spatial dynamic enhancement network. The method incorporates a low-frequency dynamic separation and enhancement mechanism in the wavelet domain low-frequency processing stage of the wavelet spatial dynamic enhancement network, and a scale-aware spatial convolution strategy in the spatial domain compensation stage after inverse wavelet reconstruction.

[0010] First, in the wavelet domain low-frequency processing stage, the low-frequency features to be enhanced are dynamically separated into smoothed low-frequency responses and low-frequency residual responses. Let the input low-frequency features be... According to the input low-frequency characteristics Generate dynamic low-pass filter weights Let the unit filter kernel be... The unit filter kernel With the dynamic low-pass filter weights Subtracting them yields the residual filter weights. ;Utilizing the dynamic low-pass filter weights For low-frequency features Local weighted aggregation is performed to obtain a smooth low-frequency response. ; and utilize the residual filter weights For low-frequency features Local weighted aggregation is performed to obtain the low-frequency residual response. Among them, smooth low-frequency response Used to characterize the brightness distribution and coarse structure information in low-frequency features, low-frequency residual response Used to characterize supplementary variation information relative to a smooth low-frequency response.

[0011] Furthermore, regarding the smooth low-frequency response and low-frequency residual response Selective fusion and gated residual updates are performed. A smooth low-frequency response is assumed. The adaptive weights are Low-frequency residual response The adaptive weights are Element-by-element multiplication is The output convolution operation is ,according to Obtain dynamic fusion low-frequency features Let the adaptive gating coefficient be... Subsequently, an adaptive gated residual update mechanism was used. Controlling the dynamic fusion of low-frequency features For the original low-frequency characteristics The correction intensity is used to obtain the enhanced low-frequency characteristics. This method allows for controlled low-frequency enhancement while preserving the original low-frequency information, thereby alleviating the problem of low-frequency brightness baseline instability caused by overall enhancement of the low-frequency sub-band.

[0012] Secondly, in the spatial domain compensation stage after inverse wavelet reconstruction, based on the network scale of the input spatial features, the corresponding spatial convolution operator is selected between fixed overlapping group convolution and ordinary spatial convolution to perform local spatial context modeling. Let the input spatial features at the s-th network scale be... The scale set of fixed-overlapping grouped convolution is... The scale set of ordinary spatial convolution is ; when At that time, for input space features Local spatial context modeling is performed using fixed-overlap grouped convolution FOGConv; when At that time, for input space features Local spatial context modeling is performed using a standard 5×5 spatial convolution. In a preferred embodiment, , .

[0013] The fixed-overlap grouped convolution FOGConv divides the input spatial features into multiple channel subgroups with fixed overlap regions according to the channel dimension, with adjacent channel subgroups sharing a portion of the input channels. After each channel subgroup undergoes a 5×5 spatial convolution, the outputs of each group are concatenated along the channel dimension, and the concatenated channel features are mixed within FOGConv using a 1×1 convolution. Let the fixed channel overlap ratio be... The fixed channel overlap ratio This represents the ratio of the number of shared channels between adjacent channel subgroups to the number of channels in a single channel subgroup; in a preferred embodiment, .

[0014] The present invention has the following beneficial effects:

[0015] First, this invention addresses the problem that coarse-grained enhancement of the low-frequency subband in existing wavelet domain low-light image enhancement methods easily leads to instability in the low-frequency brightness baseline. It proposes a low-frequency dynamic separation enhancement module (LFDSEM). This module dynamically generates low-pass filter weights based on the input low-frequency features and further constructs residual filter weights, thereby dynamically separating the low-frequency features into a smooth low-frequency response and a low-frequency residual response. Subsequently, through selective modulation and adaptive gated residual update, the low-frequency correction intensity is controlled, achieving controlled enhancement of the low-frequency brightness structure. This approach alleviates the unnatural brightness transition and low-frequency artifacts caused by overall low-frequency subband enhancement, improving the stability of low-light image brightness recovery and main structure preservation.

[0016] Secondly, to address the issue of redundant parameters and computational overhead caused by using conventional spatial convolution at all scales in the spatial domain compensation stage after inverse wavelet reconstruction, this invention proposes a scale-aware spatial convolution strategy, SASC. This strategy selects different spatial convolution operators based on the scale of the features. Fixed-overlap grouped convolution FOGConv is used in the high-resolution or detail restoration stages to reduce redundant spatial convolution computations; while ordinary 5×5 convolution is retained in the intermediate compact feature representation stage to maintain local spatial modeling capabilities. In this way, spatial reconstruction capabilities can be maintained while reducing model complexity.

[0017] This invention can effectively improve the stability of brightness recovery, structure preservation and detail expression of low-light images, while maintaining a natural visual effect and controlling the computational complexity of the model.

[0018] The significance of this invention lies in improving the visibility, clarity, and naturalness of images in complex low-light scenes, enabling the enhanced images to more realistically present the main structure and texture details.

[0019] This invention is applicable to image preprocessing scenarios such as low-light image enhancement, nighttime scene image restoration, and subsequent visual perception tasks. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the wavelet spatial dynamic enhancement network WSDEN described in this invention.

[0021] Figure 2 This is a schematic diagram comparing the structure of the fixed overlap grouped convolution FOGConv described in this invention with that of ordinary grouped convolution.

[0022] Figure 3 This is a low-light image enhancement visualization comparison image of a randomly selected image from the LOL-v1 public dataset using the method described in this invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0024] Low-light image enhancement is an image processing procedure that restores brightness, preserves structure, and enhances details in images acquired under low-light conditions. By enhancing low-light images, visibility, sharpness, and naturalness can be improved, providing more reliable image input for subsequent visual tasks such as object detection, image recognition, semantic segmentation, intelligent monitoring, and autonomous driving.

[0025] Figure 1 This is a model architecture diagram of the wavelet spatial dynamic enhancement network WSDEN described in this invention, which includes the internal structure of WSDEN's multi-scale coding and decoding structure, wavelet spatial collaborative feature enhancement unit WSCEU, wavelet domain branch module WDBM, spatial domain reconstruction feature enhancement block SRFEB, and low-frequency dynamic separation enhancement module LFDSEM. Figure 2 This is a schematic diagram of the fixed overlap grouped convolution FOGConv described in this invention, and it shows the difference between FOGConv and ordinary grouped convolution. Figure 3 This is a low-light image enhancement visualization comparison image of an image randomly selected from the LOL-v1 public dataset for this invention.

[0026] The low-light image enhancement method proposed in this invention mainly consists of the following steps: low-light image input and shallow feature extraction, wavelet domain low-frequency dynamic separation enhancement, wavelet domain high-frequency detail restoration, spatial domain compensation enhancement after inverse wavelet reconstruction, multi-scale feature fusion, and enhanced image reconstruction. Figure 1 As shown, the specific explanations for each step are as follows.

[0027] Step S1: Input a low-light image, and perform feature mapping on the low-light image through a shallow feature extraction layer to obtain shallow image features.

[0028] Input low-light image denoted as ,in, and These represent the height and width of the image, respectively, and 3 represents the number of channels in the RGB image. The low-light image is input into a shallow feature extraction layer, and initial image features are extracted through shallow convolution operations to obtain the shallow image features. .

[0029] Specifically, the shallow feature extraction layer employs Convolution is used to map the input low-light image from the image space to the feature space to obtain the input features of the subsequent wavelet spatial collaborative feature enhancement unit (WSCEU).

[0030] Step S2: Input the shallow image features or network intermediate features obtained in step S1 into the wavelet domain branch module WDBM in the wavelet spatial collaborative feature enhancement unit (WSCEU). Perform wavelet decomposition, low-frequency subband dynamic enhancement, high-frequency subband directional enhancement, and inverse wavelet reconstruction on the input features to obtain wavelet reconstructed features.

[0031] Step S2-1: Perform a discrete wavelet transform on the input features, decomposing the input features into a low-frequency sub-band and three directional high-frequency sub-bands. The low-frequency sub-band is used to characterize the overall brightness, color distribution, and coarse-scale structural information in the low-light image, while the three directional high-frequency sub-bands are used to characterize the edge, texture, and directional detail information in the low-light image.

[0032] Specifically, WDBM first performs a discrete wavelet transform (WT) on the input feature X, decomposing it into a low-frequency subband L and high-frequency subbands H, V, and D in the horizontal, vertical, and diagonal directions:

[0033]

[0034] In this invention, the low-frequency subband and high-frequency subband refer to the relative frequency subbands obtained under the same level of two-dimensional discrete wavelet decomposition. The low-frequency subband L is obtained by combining low-pass filters in the horizontal and vertical directions, mainly preserving the overall brightness, color distribution, and smooth coarse structure information in the input features that change relatively slowly in space. The high-frequency subbands H, V, and D are obtained by different combinations of low-pass and high-pass filters in the horizontal and vertical directions, mainly preserving the edge, texture, and directional detail information that changes more rapidly than the low-frequency components.

[0035] After two-dimensional discrete wavelet transform and double downsampling, the number of channels in the low-frequency sub-band and the three high-frequency sub-bands in the three directions are the same as the input feature, and the height and width are half of the height and width of the input feature, respectively.

[0036] Step S2-2: Input the low-frequency sub-band L obtained in step S2-1 into the low-frequency processing branch of WDBM. The low-frequency processing branch includes WTConv wavelet convolution and LFDSEM low-frequency dynamic separation and enhancement module; firstly, feature extraction is performed on the low-frequency sub-band through WTConv wavelet convolution to expand the effective receptive field of low-frequency features and enhance the information expression related to overall brightness, main structure and coarse-scale content, and then the obtained low-frequency features are input into the LFDSEM low-frequency dynamic separation and enhancement module.

[0037] Specifically, low-frequency subband First, the low-frequency features are processed by WTConv to obtain the features before inputting into LFDSEM. ,

[0038]

[0039] in, This represents the low-frequency features after processing by WTConv. In this embodiment, the low-frequency subband passes through two layers of WTConv sequentially. The number of input channels and output channels in each layer of WTConv is C, the kernel size is 5×5, and the number of wavelet decomposition layers is 3.

[0040] Step S2-3: The LFDSEM low-frequency dynamic separation enhancement module, based on the low-frequency characteristics obtained in step S2-2... Dynamic low-pass filter weights are generated, and these weights are used to locally weight and aggregate low-frequency features to obtain a smooth low-frequency response. Simultaneously, the unit filter kernel is subtracted from the dynamic low-pass filter weights to obtain residual filter weights, which are then used to locally weight and aggregate low-frequency features to obtain a low-frequency residual response. The smooth low-frequency response primarily corresponds to brightness distribution and coarse structure information, while the low-frequency residual response represents supplementary variation information relative to the smooth low-frequency response.

[0041] Specifically, let the low-frequency features obtained in step S2-2 be... Where B represents the number of input samples, and C represents the total number of channels for low-frequency features in the current network stage. and These represent the height and width of the low-frequency characteristics, respectively. The C channels are divided into G channel groups, with each channel group containing the following number of channels: In this embodiment, the number of channel groups is preferably set to G=8; when the number of channels C in the current network stage is not divisible by 8, the number of channel groups is set to G=1.

[0042] LFDSEM first focuses on low-frequency characteristics The global statistical information is used to generate dynamic low-pass filter weights. Specifically, for Adaptive global average pooling is performed to obtain channel statistical descriptions; then, grouped dynamic filter kernels are generated through convolutional mapping, two-dimensional batch normalization, and shape rearrangement operations; finally, Softmax normalization is performed at the k² filter kernel spatial locations corresponding to each channel group to obtain dynamic low-pass filter weights. .

[0043]

[0044] in, This indicates an adaptive global average pooling operation; This represents a channel mapping operation, used to map channel description vectors to... Each filter weight; This indicates a two-dimensional batch normalization operation; This indicates that the filter weights will be rearranged as follows: Groups, each group is sized as follows Dynamic filter kernel; Indicates the dimension of the convolution kernel space Normalize the above; This represents the dynamically generated low-pass filter weights. It should be noted that this formula includes a function variable. and These represent channel mapping operation and two-dimensional batch normalization operation, respectively; B and C without function variables represent the number of input samples and the total number of channels for low-frequency features, respectively, and their meanings are different.

[0045] The dynamic filtering kernel generates a convolutional layer with a kernel size of 1×1, a stride of 1, and no bias term. Its input channels are C, and its output channels are Gk². In this embodiment, the dynamic filtering kernel size is preferably set to k=3. When G=8 and k=3, the number of output channels generated by the dynamic filtering kernel is... .

[0046] To obtain a low-frequency residual response that complements the smooth low-frequency response, LFDSEM uses a unit filter kernel. With dynamic low-pass filter weights Subtracting them yields the residual filter weights. .

[0047]

[0048] Where I represents the unit filter kernel, whose spatial size is the same as that of the unit filter kernel. The same applies, with a value of 1 at the center position and 0 at all other positions. When the unit filter kernel I acts on the input low-frequency features, it preserves the original low-frequency features, and the dynamic low-pass filter weights... This is used to obtain a local smoothing response for low-frequency characteristics of the input. Therefore, by... The residual filter weights constructed by subtracting from the unit filter kernel I. This method can extract supplementary variation information from the original low-frequency features that were not represented by the smoothed low-frequency response. When the local low-frequency features are relatively smooth, the resulting low-frequency residual response is relatively small; when there are brightness or structural changes relative to the smoothed baseline in a local region, the resulting low-frequency residual response increases accordingly. In this way, the input low-frequency features are separated into complementary smoothed low-frequency responses and low-frequency residual responses, providing different types of low-frequency information for subsequent selective fusion.

[0049] Based on the aforementioned channel grouping relationship, the channel dimensions of low-frequency features are rearranged from C to... Utilizing the dynamic low-pass filter weights corresponding to each channel group. and residual filter weights Low-frequency features Local weighted aggregation is performed to obtain a smooth low-frequency response. and low-frequency residual response .

[0050]

[0051] in, Indicates the sample index; Indicates the channel group index; This represents the channel index within the g-th channel group; , These represent the longitudinal and lateral spatial locations of the low-frequency features, respectively. Indicates the spatial offset position within the local filter kernel. This represents a local filtering neighborhood of size k×k.

[0052] In this embodiment, according to the aforementioned dynamic filter kernel settings, k=3, therefore Before performing local weighted aggregation, low-frequency features are filled with a reflection width of 1. The step size and dilation rate of the local weighted aggregation are both 1, making... and Space dimensions and Maintain consistency.

[0053] For the same channel group g in the same sample b, the channels within that channel group Each channel shares the same dynamic low-pass filter kernel and corresponding residual filter kernel. The dynamic low-pass filter kernel is composed of weights corresponding to each spatial offset position. The residual filter kernel is composed of weights corresponding to each spatial offset position. The filter kernel is shared across different spatial locations of the sample.

[0054] It should be noted that the low-frequency residual response Unlike the high-frequency subbands in the horizontal, vertical, and diagonal directions obtained directly by discrete wavelet transform, it is low-frequency residual information obtained by dynamic residual filtering within the low-frequency processing branch.

[0055] Step S2-4: Selectively modulate the smoothed low-frequency response and low-frequency residual response obtained in step S2-3 to generate dynamically fused low-frequency features; then, control the correction intensity of the dynamically fused low-frequency features on the original low-frequency features through a sample adaptive gated residual update mechanism to obtain enhanced low-frequency features. In this way, controlled correction is performed while retaining the original low-frequency information, thereby establishing a more stable low-frequency brightness baseline.

[0056] Specifically, in obtaining a smooth low-frequency response and low-frequency residual response Then, LFDSEM first aggregates the two and generates a global description vector through adaptive global average pooling and convolutional mapping. :

[0057]

[0058] in, This represents a convolution mapping operation with a kernel size of 1×1 and a stride of 1. Its input channel number is C, and the output channel number is the larger of half of C (rounded down) and 32, denoted as d.

[0059] Subsequently, the global descriptive feature Z is input into two independent weight generation branches. Both weight generation branches use convolutional layers with a kernel size of 1×1 and a stride of 1, with d input channels and C output channels. The outputs of the two branches are concatenated along the channel dimension to form 2C channels, and Softmax normalization is performed on these 2C channels. Then, the normalized 2C channel weights are split into two groups along the channel dimension, each containing C channels, to obtain adaptive weights with smooth low-frequency responses. Adaptive weights for low-frequency residual responses :

[0060]

[0061] in, and These represent the mapping functions used to generate the corresponding response weights. This indicates that the outputs of the two mapping branches are concatenated along the channel dimension. Based on the adaptive weights, the smoothed low-frequency response and the low-frequency residual response are weighted and fused to obtain the dynamically fused low-frequency features. :

[0062]

[0063] in, This represents element-wise multiplication. This indicates the output convolution operation. In this embodiment, the output convolution uses a convolutional layer with a kernel size of 1×1 and a stride of 1, and both its input and output channels are C.

[0064] To avoid over- or under-correction caused by low-frequency correction with a fixed intensity, LFDSEM generates sample adaptive gating coefficients. This is used to control the low-frequency correction intensity. The gate coefficient... Due to low frequency characteristics generate:

[0065]

[0066] in, This represents the gated generation function. This represents the Sigmoid function. Used to constrain the maximum update magnitude. The gated generation function. The sequence includes adaptive global average pooling, a first 1×1 gated convolutional layer, a ReLU activation function, and a second 1×1 gated convolutional layer. The first gated convolutional layer has C input channels and the number of output channels is the larger of the floor function of C / 8 and 1; this number of output channels is denoted as... The number of input channels in the second gated convolutional layer is The number of output channels is 1. The resulting gating coefficient... The size is B×1×1×1, meaning each input sample corresponds to one gating coefficient, and the application is broadcast across the channel and spatial dimensions of the low-frequency features. In this embodiment, the maximum update amplitude... Preferred settings .

[0067] Subsequently, the enhanced low-frequency features are obtained through gated residual updates. As shown in equation (10):

[0068]

[0069] A smaller gating response indicates that more original low-frequency information is retained, while a larger gating response indicates that more dynamically fused correction information is introduced. In this way, LFDSEM can perform controlled correction while preserving the original low-frequency foundation, thereby establishing a more stable low-frequency luminance baseline.

[0070] Step S2-5: Input the high-frequency subbands H, V, and D obtained in step S2-1 into the high-frequency processing branch of WDBM, and perform lightweight convolution enhancement on the three high-frequency subbands; at the same time, perform directional convolution processing on the enhanced low-frequency features obtained in step S2-4, extract the low-frequency supplementary features corresponding to the H, V, and D directions respectively, and splice and fuse the low-frequency supplementary features with the high-frequency subbands of the corresponding directions to obtain the fused directional high-frequency features.

[0071] Step S2-6: Input the fused high-frequency features obtained in step S2-5 into the state space modeling module to further model the dependencies between high-frequency details in different directions, and obtain the enhanced high-frequency sub-band features.

[0072] Step S2-7: Combine the enhanced low-frequency features obtained in step S2-4 and the enhanced high-frequency sub-band features obtained in step S2-6, and perform inverse discrete wavelet transform to reconstruct spatial domain features, thus obtaining wavelet reconstructed features.

[0073] Specifically, the enhanced low-frequency characteristics With enhanced high-frequency subband and The features are combined and inverse discrete wavelet transform (IWT) is performed to obtain the wavelet reconstructed features. As shown in equation (11):

[0074]

[0075] in, This represents the spatial domain features reconstructed jointly by enhancing low-frequency features and enhancing high-frequency features.

[0076] Step S3: Input the wavelet reconstruction features obtained in step S2 into the spatial domain reconstruction feature enhancement block SRFEB in the wavelet spatial collaborative feature enhancement unit (WSCEU) to perform global context modeling, local spatial context enhancement, and channel interaction enhancement on the wavelet reconstructed spatial domain features to obtain spatial domain enhanced features.

[0077] Step S3-1: Input the wavelet reconstruction features obtained in step S2 into the spatial domain reconstruction feature enhancement block SRFEB, perform global context modeling and local spatial context enhancement on the wavelet reconstructed spatial domain features, and obtain spatial residual enhancement features.

[0078] Step S3-1-1: Normalize the wavelet reconstruction features of the input SRFEB and input them into the Fast Fourier Convolution Residual Block FFCResnetBlock to obtain spatially enhanced intermediate features; wherein, the Fast Fourier Convolution Residual Block introduces global context information through frequency domain transformation to expand the effective receptive field of spatial domain features and enhance the global structural expression ability.

[0079] Step S3-1-2: Input the spatially enhanced intermediate features obtained in step S3-1-1 into the scale-aware spatial convolution strategy SASC for local spatial context modeling. The SASC selects different spatial convolution operators according to the network scale of the features: in the high-resolution stage or detail recovery stage, fixed-overlap grouped convolution FOGConv is used; in the intermediate compact feature expression stage, ordinary 5×5 convolution is retained to maintain the local spatial modeling capability.

[0080] Specifically, for the first Input features at each stage ,

[0081]

[0082] in, Represents the input features of the s-th network stage The scale-aware spatial convolution operator used is... This represents a fixed-overlapping grouped convolution. This represents a standard 5×5 convolution; , indicating the set of stages using FOGConv; , representing the set of stages using ordinary 5×5 convolutions. This strategy reduces redundant spatial computation in the high-resolution stage or detail recovery stage, while maintaining local spatial modeling capabilities in the intermediate compact feature representation stage.

[0083] Step S3-1-3: FOGConv divides the input features into multiple channel subgroups with fixed overlap regions according to the channel dimension. Adjacent channel subgroups share some input channels. After performing a 5×5 spatial convolution on each channel subgroup, the outputs of each group are concatenated along the channel dimension. In a preferred embodiment, the fixed channel overlap ratio is set to 0.25. In this way, FOGConv can reduce redundant computation of local spatial convolutions at high resolution while enhancing information interaction between adjacent channel groups.

[0084] Specifically, as shown in Figure 2, ordinary grouped convolution divides the input features into multiple non-overlapping channel groups according to the channel dimension. Each channel group undergoes spatial convolution before the output is concatenated. Unlike ordinary grouped convolution, FOGConv sets a fixed overlap region between adjacent channel subgroups, allowing adjacent channel groups to share some input channels. Given input features, FOGConv first divides them into multiple channel subgroups with fixed overlap regions; then, it performs a 5×5 spatial convolution on each channel subgroup; subsequently, it concatenates the outputs of each group along the channel dimension; finally, within FOGConv, it mixes the concatenated channel features using a 1×1 convolution to obtain the FOGConv output features. The fixed channel overlap ratio is used in this process. This indicates the proportion of shared channels between adjacent channel subgroups, preferably set to 0.25.

[0085] Step S3-1-4: When FOGConv is used in the current stage, the local spatial context enhancement features are output by FOGConv; when ordinary 5×5 convolution is used in the current stage, the local spatial context enhancement features are output by ordinary 5×5 convolution.

[0086] Step S3-1-5: Perform residual fusion between the local spatial context enhancement features obtained in step S3-1-4 and the wavelet reconstruction features before inputting SRFEB to obtain spatial residual enhancement features.

[0087] Step S3-2: Input the spatial residual enhancement features obtained in step S3-1 into the channel interaction enhancement stage in SRFEB to enhance the information interaction of the spatial residual enhancement features in the channel dimension. Specifically, first, the global context representation is enhanced by normalization and fast Fourier convolution residual blocks, then the channel interaction is enhanced by 1×1 point convolution and gating mechanism, and finally, residual fusion is performed with the spatial residual enhancement features before inputting to this stage to obtain the spatial domain enhancement features.

[0088] Step S4: In the wavelet spatial dynamic enhancement network WSDEN, the WSCEU constructed in steps S2 and S3 is repeated at different scales, and the multi-scale features are enhanced and fused by combining downsampling and upsampling layers. Finally, the enhancement residual is predicted by the output reconstruction layer, and the enhancement residual is added to the input low-light image to generate the final enhanced image.

[0089] Specifically, in the encoding stage, the network progressively reduces the feature space resolution and expands the receptive field through downsampling layers; in the decoding stage, the network progressively restores the spatial resolution through upsampling layers and fuses multi-scale features using skip connections. WSDEN includes two types of WSCEU: WSCEU-O, which indicates that its SRFEB uses FOGConv; and WSCEU-C, which indicates that its SRFEB uses ordinary 5×5 convolutions. Preferably, the first, second, and fifth stages use WSCEU-O, and the third and fourth stages use WSCEU-C.

[0090] The features after multi-scale enhancement and fusion are input into the output reconstruction layer, and the output reconstruction layer predicts the enhancement residuals. Ultimately, this will enhance the residuals. With input low-light image Add them together to get the enhanced image. .

[0091] In this embodiment, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Learning Perceptual Image Patch Similarity (LPIPS), number of parameters, and Floating-Point Operations (FLOPs) are used as evaluation metrics. Higher PSNR and SSIM indicate better reconstruction quality and structural similarity of the enhanced image; lower LPIPS indicate less perceptual distortion; and lower parameter count and FLOPs indicate lower model complexity. On the LOL-v1 public dataset, the base network without the Low-Frequency Dynamic Separation Enhancement Module (LFDSEM) and the Scale-Aware Spatial Convolution Strategy (SASC) has PSNR, SSIM, and LPIPS of 23.60 dB, 0.8496, and 0.0648, respectively. With only LFDSEM, PSNR, SSIM, and LPIPS are 24.21 dB, 0.8558, and 0.0634, respectively, demonstrating that LFDSEM effectively improves the brightness recovery quality and structural preservation ability of low-light images while reducing perceptual distortion. After simultaneously employing LFDSEM and SASC, the PSNR, SSIM, and LPIPS of the complete WSDEN method described in this invention are 23.81 dB, 0.8523, and 0.0641, respectively, still superior to the basic network. The results show that LFDSEM can establish a more stable low-frequency brightness baseline by dynamically separating and smoothing the low-frequency response and the low-frequency residual response, thereby improving the quality of low-light image enhancement. Furthermore, the number of network parameters and FLOPs using only LFDSEM are 1.26M and 11.34G, respectively; after using the complete WSDEN, the number of parameters and FLOPs are reduced to 1.21M and 10.25G, respectively. This demonstrates that SASC can reduce the number of model parameters and computational overhead while maintaining superior enhancement performance compared to the basic network. Therefore, the method described in this invention achieves a good balance between brightness recovery stability, structure preservation capability, and model computational efficiency.

[0092] The visualization comparison results using the method described in this invention on the LOL-v1 public dataset are as follows: Figure 3 As shown. Figure 3 The image demonstrates the input low-light image, a reference normal-light image, the CWNet low-light image enhancement method, and the enhancement results of the method of this invention. Figure 3 As can be seen, the method of the present invention can improve the overall brightness while maintaining a more natural color distribution and alleviate the unnatural brightness transition caused by the instability of the low-frequency brightness baseline.

Claims

1. A low-light image enhancement method based on wavelet spatial dynamic enhancement network, characterized in that, include: In the wavelet domain low-frequency processing stage of the wavelet spatial dynamic enhancement network, the low-frequency features to be enhanced are dynamically separated into smooth low-frequency response and low-frequency residual response. The separated smooth low-frequency response and low-frequency residual response are selectively fused and gated residual updates are performed to obtain the enhanced low-frequency features. In the spatial domain compensation stage of the wavelet spatial dynamic enhancement network, based on the network scale of the input spatial features, the corresponding spatial convolution operator is selected between fixed overlapping group convolution and ordinary spatial convolution to perform local spatial context modeling, thereby obtaining local spatial context enhancement features.

2. The low-light image enhancement method based on wavelet spatial dynamic enhancement network according to claim 1, characterized in that, The dynamic separation, selective fusion, and gated residual update of the smoothed low-frequency response and low-frequency residual response include: Let the input low-frequency characteristics be According to the input low-frequency characteristics Generate dynamic low-pass filter weights ; Let the unit filter kernel be The unit filter kernel With the dynamic low-pass filter weights Subtracting them yields the residual filter weights. ; Using the dynamic low-pass filter weights For low-frequency features Local weighted aggregation is performed to obtain a smooth low-frequency response. ; and utilize the residual filter weights For low-frequency features Local weighted aggregation is performed to obtain the low-frequency residual response. ; Assume a smooth low-frequency response The adaptive weights are Low-frequency residual response The adaptive weights are Element-by-element multiplication is The output convolution operation is ,according to Obtain dynamic fusion low-frequency features ; Let the adaptive gating coefficient be Subsequently, an adaptive gated residual update mechanism was used. Controlling the dynamic fusion of low-frequency features For the original low-frequency characteristics The correction intensity is used to obtain the enhanced low-frequency characteristics. .

3. The low-light image enhancement method based on wavelet spatial dynamic enhancement network according to claim 1, characterized in that, The step of selecting the corresponding spatial convolution operator between fixed-overlapping group convolution and ordinary spatial convolution for local spatial context modeling based on the network scale of the input spatial features includes: Let the input spatial features of the s-th network scale be... The scale set of fixed-overlapping grouped convolution is... The scale set of ordinary spatial convolution is ; when At that time, for input space features Local spatial context modeling is performed using fixed-overlap grouped convolution FOGConv; when At that time, for input space features Local spatial context modeling is performed using a standard 5×5 spatial convolution. The fixed-overlap grouped convolution FOGConv divides the input spatial features into multiple channel subgroups with fixed overlap regions along the channel dimension, with adjacent subgroups sharing some input channels. After each channel subgroup undergoes a 5×5 spatial convolution, the outputs of each group are concatenated along the channel dimension, and the concatenated channel features are mixed within FOGConv using a 1×1 convolution. The fixed channel overlap ratio is set to... The fixed channel overlap ratio This indicates the proportion of shared channels between adjacent channel subgroups to the total number of channels in a single channel subgroup, and In a preferred embodiment, , .

Images