A low-light image enhancement method based on frequency domain decoupling and state space model

Abstract

This invention relates to the field of computer vision and discloses a low-light image enhancement method based on frequency domain decoupling and a state-space model. The method includes the following steps: a low-light image is preprocessed to obtain a feature map, which is then input into a pre-trained enhancement network. The network contains multiple coding units, latent layers, and corresponding decoding units. The feature map is first enhanced by the WMH module of the first coding unit, then downsampled to progressively reduce resolution and increase the number of channels until the last coding unit outputs a deep feature map to the latent layer. The latent layer performs frequency domain adjustment to obtain an optimized feature map. The optimized feature map is first upsampled to restore resolution, then weighted and fused with the enhanced feature map output by the corresponding coding unit via skip connections. After further enhancement by the WMH module, it is fed into the next decoding unit. Finally, the decoding unit outputs the enhanced image through convolutional mapping. This method provides an efficient, robust, and more universal technical solution.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and in particular to a low-light image enhancement method based on frequency domain decoupling and a state-space model. Background Technology

[0002] In the field of computer vision, images serve as the core information carrier, and their quality directly affects the accuracy of downstream tasks such as security monitoring, autonomous driving, and facial recognition. However, images acquired in reality often suffer from quality degradation due to factors such as insufficient lighting (low-light scenes) and atmospheric scattering (foggy scenes), resulting in problems such as low brightness, poor contrast, loss of detail, and noise interference, which severely restricts the progress of subsequent tasks.

[0003] Low-light image enhancement technology improves image visual quality through algorithmic optimization and has become a research hotspot in the field of image processing. Traditional methods, such as histogram equalization, are prone to color imbalance and noise amplification. Retinex series methods are limited by their decomposition capabilities, resulting in limited enhancement effects. Among deep learning-based methods, CNN-type models are limited by their local receptive fields and struggle to capture long-distance dependencies. While Transformer-type models can model global information, their computational cost is too high to meet the needs of real-time applications.

[0004] In recent years, the Mamba architecture, as a novel deep learning architecture based on the state space model, has shown great potential in visual tasks due to its efficient long sequence processing capabilities and hardware friendliness. Wavelet transform has been widely used in the field of image enhancement because it can effectively decouple high and low frequency information in images. However, existing algorithms that combine wavelet transform lack targeted optimization design for high and low frequency information, and most of them are only applicable to single degradation scenarios, making it difficult to achieve cross-scenario general enhancement. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a low-light image enhancement method based on frequency domain decoupling and a state-space model, which offers an efficient, robust, and more versatile technical solution.

[0006] This invention provides a low-light image enhancement method based on frequency domain decoupling and a state-space model, comprising the following steps: Acquire a low-light input image and preprocess it to obtain a feature map; The feature map is input into a pre-trained augmentation network, which includes multiple encoding units, a latent layer and multiple decoding units connected in sequence, with each encoding unit corresponding to a decoding unit. The feature map is first input into the first layer coding unit, and then enhanced by the WMH module to obtain an enhanced feature map. The enhanced feature map is then downsampled between adjacent coding units to gradually reduce the feature map resolution and increase the number of channels before being input into the next layer coding unit. The above operation is repeated until the deep feature map output by the last coding unit is downsampled and input into the latent layer. The latent layer performs frequency domain adjustment on the input deep feature map to obtain an optimized feature map; Before being input into the first decoding unit, the optimized feature map is upsampled to restore the resolution. The resulting upsampled feature map is then weighted and fused with the enhanced feature map output by the corresponding encoding unit through skip connections. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The above upsampling operation is repeated until the feature map output by the last decoding unit is convolved to obtain the enhanced image. Specifically, the following sub-steps are performed in the WMH module of each encoding unit and each decoding unit: The input feature map is decomposed into low-frequency sub-bands and high-frequency sub-bands using discrete wavelet transform; For the low-frequency sub-band, a hybrid CNN-state space model module is used to extract local features and model the global context to obtain enhanced low-frequency features; A parallel spatial-channel dual attention module is used to enhance texture and suppress noise in the high-frequency subband to obtain enhanced high-frequency features. The enhanced low-frequency and high-frequency features are reconstructed into an enhanced feature map through inverse discrete wavelet transform and then output.

[0007] Specifically, the WMH module includes: The discrete wavelet transform unit is used to decompose the input feature map into a low-frequency subband and three high-frequency subbands in three directions; The high-frequency processing submodule is used to perform spatial-channel dual attention enhancement on the three high-frequency subbands to obtain the enhanced high-frequency features. The low-frequency processing submodule is used to perform CNN local feature extraction and bidirectional state-space model global modeling on the low-frequency subband to obtain enhanced low-frequency features. The inverse discrete wavelet transform unit is used to reconstruct the enhanced low-frequency subband and high-frequency subband into enhanced feature maps and output them.

[0008] Specifically, the high-frequency processing submodule employs parallel spatial attention and channel attention mechanisms: The channel attention branch extracts channel descriptors through global average pooling and global max pooling, and generates channel weights through shared MLP; The spatial attention branch generates spatial weights by aggregating channel information through convolution; The enhanced high-frequency features are obtained by adding the channel-weighted features and spatial-weighted features element by element, followed by convolution and normalization.

[0009] Specifically, the low-frequency processing submodule includes: The CNN local feature extraction layer is used to extract local structural information of low-frequency subbands and output a two-dimensional low-frequency feature map. The serialization unit is used to flatten a two-dimensional low-frequency feature map into a one-dimensional sequence; A bidirectional state-space model unit is used to perform forward and backward scanning of a one-dimensional sequence, fuse bidirectional context information, and output an enhanced one-dimensional sequence. The deserialization unit is used to restore the enhanced one-dimensional sequence to a two-dimensional feature map. Specifically, the bidirectional state-space model unit adopts the Mamba structure, and its state update process is represented as follows: , , Where, x t Given the input sequence, h t In the hidden state, y t The output is A, B, C, and D, which are learnable parameters.

[0010] Specifically, the number of multiple coding units is the same as the number of multiple decoding units. The skip connection of the j-th layer decoding unit receives the feature map output by the j-th layer coding unit, where j=1,2,...,N, N is the total number of coding units, and the first layer consists of the first input coding unit and the last output decoding unit.

[0011] Specifically, the latent layer of the enhanced network is equipped with a frequency domain processing module, which is used to adjust the frequency domain of the deep feature map output by the encoder and optimize the expression of texture information.

[0012] Specifically, the discrete wavelet transform and the inverse discrete wavelet transform satisfy the following relationship: , , Where LL represents the low-frequency subband, LH, HL, and HH represent high-frequency subbands in different directions, and LL', LH', HL', and HH' represent the enhanced subband characteristics, respectively. For the input feature map, The reconstructed feature map.

[0013] Specifically, the augmentation network is trained using a composite loss function, which is expressed as follows: , in, Used to optimize pixel-level precision Used to maintain image structural similarity Used to suppress image noise and artifacts, 0.95 and 0.05 are weighting coefficients.

[0014] A low-light image enhancement system based on frequency domain decoupling and a state-space model includes: The image acquisition module is used to acquire low-light input images and preprocess the images to obtain feature maps; The enhancement processing module is used to load pre-trained enhancement networks; The enhanced processing module includes: Multiple coding processing units are embedded with WMH modules. Each coding processing unit is used to receive feature maps after downsampling to reduce resolution and increase the number of channels. The feature maps are then enhanced by the WMH module to obtain enhanced feature maps, which are then downsampled and input into the next coding unit. The deep feature maps output by the last coding unit are downsampled and input into the latent layer. The latent layer processing unit is used to perform frequency domain adjustment on the input deep feature map to obtain an optimized feature map, which is then upsampled and input to the first decoding unit. Multiple decoding processing units are embedded with WMH modules. Each decoding unit is used to receive the upsampled feature map obtained by upsampling to restore the resolution, and then perform skip connection weighted fusion with the enhanced feature map output by the corresponding encoding unit. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The feature map output by the last decoding unit is convolved to obtain the enhanced image. The WMH modules all include: Discrete wavelet transform unit is used to decompose the input feature map into low-frequency subband and high-frequency subband; The low-frequency processing submodule is used to perform local feature extraction and global context modeling on the low-frequency subband using a hybrid CNN-state space model module to obtain enhanced low-frequency features. The high-frequency processing submodule is used to perform texture enhancement and noise suppression on the high-frequency subband using a parallel spatial-channel dual attention module to obtain enhanced high-frequency features. The inverse discrete wavelet transform unit is used to reconstruct the enhanced low-frequency features and high-frequency features into an enhanced feature map and output it. The output module is used to output the enhanced image.

[0015] The technical solution provided by this invention has the following advantages compared with the prior art: It utilizes discrete wavelet transform to decouple image features in the frequency domain, thereby achieving separate processing and collaborative optimization of low-frequency components reflecting image structure and high-frequency components containing texture details and noise. Its advantage lies in the fact that by introducing discrete wavelet transform to achieve frequency domain decoupling of image features, complex degraded images are decomposed into low-frequency components containing the main structure and high-frequency components containing texture details and noise. This provides a clear physical basis for subsequent targeted optimization and effectively solves the contradiction between detail blurring and noise amplification that easily occurs when traditional methods process in a unified spatial domain. For the low-frequency components, it innovatively integrates the local feature extraction capability of CNNs with the global context modeling capability of state-space models. This hybrid architecture not only compensates for the shortcomings of traditional CNNs, such as limited receptive field and difficulty in capturing long-distance illumination dependencies, but also overcomes the bottleneck of high computational overhead and difficulty in meeting real-time requirements of Transformer models, achieving accurate restoration of the overall brightness and contrast of the image. For high-frequency components, a parallel spatial-channel dual attention mechanism is designed, which can adaptively enhance effective texture edges and suppress noise in different dimensions. This solves the problems of existing methods lacking selectivity, easily amplifying noise or losing details when processing high-frequency information, and significantly improves the detail clarity and purity of the enhanced image. This method combines the flexibility of frequency domain decoupling with the hierarchical structure of the encoding and decoding network. Its core processing mechanism is not limited to low-light scenes, but also has general enhancement potential for other frequency domain degradation scenes such as fog. Therefore, it provides an efficient, robust and more universal technical solution for solving the problem of image quality degradation in complex environments. Attached Figure Description

[0016] Figure 1 A schematic diagram of the main network structure of a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a DAHFP module for a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a low-frequency processing submodule of a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention. Figure 4 A schematic diagram of the inverse discrete wavelet transform of a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of the WMH module in a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention; Figure 6A flowchart of a low-light image enhancement method based on frequency domain decoupling and state space model provided in an embodiment of the present invention; Figure 7 A flowchart of the sub-steps executed by the WMH module provided in an embodiment of the present invention. Detailed Implementation

[0017] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0018] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0019] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0020] Figure 1 This is a schematic diagram of the main network WalMHNet, which is provided in an embodiment of the present invention for a low-light image enhancement method based on frequency domain decoupling and a state-space model. Figure 2 This is a schematic diagram of the DAHFP module, which is a low-light image enhancement method based on frequency domain decoupling and state space model, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the MGCM module, which is a low-light image enhancement method based on frequency domain decoupling and state-space model, according to an embodiment of the present invention. Figure 4 This is a wavelet decomposition diagram illustrating a low-light image enhancement method based on frequency domain decoupling and a state-space model, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the WMH core module of a low-light image enhancement method based on frequency domain decoupling and state-space model provided in an embodiment of the present invention. Figure 6 This is a flowchart illustrating a low-light image enhancement method based on frequency domain decoupling and a state-space model, provided in an embodiment of the present invention. Figure 7 A flowchart of the sub-steps executed by the WMH module provided in an embodiment of the present invention.

[0021] like Figure 1 and Figure 6 As shown, this embodiment of the invention provides a low-light image enhancement method based on frequency domain decoupling and a state-space model, including the following steps: acquiring a low-light input image and preprocessing it to obtain a feature map; inputting the feature map into a pre-trained enhancement network, the enhancement network including multiple encoding units, a latent layer, and multiple decoding units connected in sequence, with each encoding unit corresponding to a decoding unit; the feature map is first input into the first layer encoding unit, and is first enhanced by the WMH module to obtain an enhanced feature map; the enhanced feature map is then input into the next layer encoding unit after downsampling to gradually reduce the feature map resolution and increase the number of channels between adjacent encoding units, and the above operation is repeated until the deep feature map output by the last encoding unit is input into the latent layer after downsampling; the latent layer performs frequency domain adjustment on the input deep feature map to obtain an optimized feature map; the optimized feature map is upsampled to restore resolution before being input into the first decoding unit. The obtained upsampled feature map is then fused with the enhanced feature map output by the corresponding encoding unit through skip connections and weighted fusion. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The above upsampling operation is repeated until the feature map output by the last decoding unit is convolved to obtain the enhanced image. In each encoding unit and each decoding unit's WMH module, the following sub-steps are performed: the input feature map is decomposed into low-frequency sub-bands and high-frequency sub-bands through discrete wavelet transform; the low-frequency sub-bands are subjected to local feature extraction and global context modeling using a hybrid CNN-state space model module to obtain enhanced low-frequency features; the high-frequency sub-bands are subjected to texture enhancement and noise suppression using a parallel spatial-channel dual attention module to obtain enhanced high-frequency features; the enhanced low-frequency features and high-frequency features are reconstructed into an enhanced feature map through inverse discrete wavelet transform and output.

[0022] Specifically, the downsampling operation uses pixel rearrangement downsampling, which involves first adjusting the number of feature channels through a convolutional layer, then reducing the spatial resolution of the feature map through PixelUnshuffle, and rearranging the spatial information to the channel dimension. The upsampling operation uses pixel rearrangement upsampling, which involves first adjusting the number of feature channels through a convolutional layer, then rearranging the channel information of the feature map to the spatial dimension through PixelShuffle, thereby restoring the feature map to its original resolution.

[0023] `PixelUnshuffle` and `PixelShuffle` are PyTorch library functions. `PixelShuffle` is a commonly used rearrangement operation in deep learning models such as image super-resolution and generative adversarial networks. `PixelUnshuffle` is the inverse operation of `PixelShuffle`. It compresses information from the spatial dimensions (height and width) to the channel dimension, achieving a transformation from width to depth.

[0024] Specifically, the WMH module includes: a discrete wavelet transform unit, used to decompose the input feature map into a low-frequency sub-band and three high-frequency sub-bands in three directions; a high-frequency processing sub-module, used to jointly perform spatial-channel dual attention enhancement on the three high-frequency sub-bands to obtain enhanced high-frequency features; a low-frequency processing sub-module, used to perform CNN local feature extraction and bidirectional state-space model global modeling on the low-frequency sub-bands to obtain enhanced low-frequency features; and an inverse discrete wavelet transform unit, used to reconstruct the enhanced low-frequency sub-band and high-frequency sub-band into an enhanced feature map and output it.

[0025] Specifically, the high-frequency processing submodule adopts a parallel spatial attention and channel attention mechanism: the channel attention branch extracts channel descriptors through global average pooling and global max pooling, and generates channel weights through shared MLP; the spatial attention branch aggregates channel information through convolution to generate spatial weights; the channel weighted features and spatial weighted features are added element by element, and then convolution and normalization are performed to obtain the enhanced high-frequency features.

[0026] Specifically, the low-frequency processing submodule includes: a CNN local feature extraction layer, used to extract local structural information of low-frequency subbands and output a two-dimensional low-frequency feature map; a serialization unit, used to flatten the two-dimensional low-frequency feature map into a one-dimensional sequence; a bidirectional state-space model unit, used to perform forward and backward scanning on the one-dimensional sequence, fuse bidirectional contextual information, and output an enhanced one-dimensional sequence; and a deserialization unit, used to restore the enhanced one-dimensional sequence into a two-dimensional feature map. Specifically, the bidirectional state-space model unit adopts the Mamba structure, and its state update process is represented as follows: , , Where, x t Given the input sequence, h t In the hidden state, y t The output is A, B, C, and D, which are learnable parameters.

[0027] Specifically, the number of multiple coding units is the same as the number of multiple decoding units. The skip connection of the j-th layer decoding unit receives the feature map output by the j-th layer coding unit, where j=1,2,...,N, N is the total number of coding units, and the first layer consists of the first input coding unit and the last output decoding unit.

[0028] Specifically, the latent layer of the enhanced network is equipped with a frequency domain processing module, which is used to adjust the frequency domain of the deep feature map output by the encoder and optimize the expression of texture information.

[0029] Specifically, the discrete wavelet transform and the inverse discrete wavelet transform satisfy the following relationship: , , Where LL represents the low-frequency subband, LH, HL, and HH represent high-frequency subbands in different directions, and LL', LH', HL', and HH' represent the enhanced subband characteristics, respectively. For the input feature map, The reconstructed feature map.

[0030] Specifically, the augmentation network is trained using a composite loss function, which is expressed as follows: , in, Used to optimize pixel-level precision Used to maintain image structural similarity Used to suppress image noise and artifacts, 0.95 and 0.05 are weighting coefficients.

[0031] A low-light image enhancement system based on frequency domain decoupling and a state-space model includes: an image acquisition module for acquiring a low-light input image and preprocessing the image to obtain a feature map; an enhancement processing module for loading a pre-trained enhancement network; the enhancement processing module includes: multiple encoding processing units, each embedding a WMH module, each encoding processing unit receiving a feature map after downsampling to reduce resolution and increase the number of channels, the feature map being enhanced by the WMH module to obtain an enhanced feature map, which is then downsampled and input into the next encoding unit, and the deep feature map output by the last encoding unit being downsampled and input into the latent layer; a latent layer processing unit for frequency domain adjustment of the input deep feature map to obtain an optimized feature map, which is then upsampled and input into the first decoding unit; and multiple decoding processing units, each embedding a WMH module, each decoding unit receiving an upsampled feature map obtained by upsampling to restore resolution. The input feature map is first upsampled and then weighted and fused with the enhanced feature map output by the corresponding coding unit through skip connections. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The feature map output by the last decoding unit is convolved to obtain the enhanced image. The WMH module includes: a discrete wavelet transform unit, which decomposes the input feature map into low-frequency sub-bands and high-frequency sub-bands; a low-frequency processing sub-module, which performs local feature extraction and global context modeling on the low-frequency sub-bands using a hybrid CNN-state space model module to obtain enhanced low-frequency features; a high-frequency processing sub-module, which performs texture enhancement and noise suppression on the high-frequency sub-bands using a parallel spatial-channel dual attention module to obtain enhanced high-frequency features; an inverse discrete wavelet transform unit, which reconstructs the enhanced low-frequency features and high-frequency features into an enhanced feature map and outputs it; and an output module, which outputs the enhanced image.

[0032] A specific embodiment includes: an input layer, an encoding module, a decoding module, a WMH module, a latent layer Fast Fourier Transform module, a skip connection module, and an output layer. The input layer receives low-light images and performs preprocessing such as resizing and normalization to ensure the consistency and validity of the input data. The encoding module consists of multiple encoding units, each including a downsampling layer and a WMH module. The downsampling layer reduces the feature map resolution and increases the number of channels through convolution and pooling operations to extract abstract features. The WMH module performs frequency domain enhancement processing on the encoded feature map. The latent layer incorporates a Fast Fourier Transform module to perform Fourier adjustment on the deep features output by the encoding module, optimizing texture information and improving feature representation capabilities. The decoding module is symmetrical to the encoding module and consists of multiple decoding units. Each decoding unit includes an upsampling layer, a skip connection fusion layer, and a WMH module. The upsampling layer restores the feature map resolution through transposed convolution or interpolation operations. The skip connection fusion layer fuses the high-resolution spatial detail features of the corresponding encoding unit with the deep semantic features of the decoding unit using learnable weights. The WMH module performs frequency domain optimization on the fused feature map. Finally, the feature map output by the decoding module is mapped to the final enhanced image through a convolutional layer.

[0033] The WMH module is the core of this invention, such as Figure 5 As shown, frequency domain processing is performed on the feature map during the encoding and decoding process, including wavelet transform, high-frequency processing module, low-frequency processing module, and inverse wavelet transform. First, a two-dimensional discrete wavelet transform is performed on the input feature map, decomposing it into one low-frequency sub-band and three high-frequency sub-bands. The low-frequency sub-band contains the main structure and global brightness information of the image, reflecting the overall features of the image; the high-frequency sub-bands correspond to high-frequency information in the horizontal-vertical, vertical-horizontal, and diagonal directions, respectively, containing texture details, edge features, and noise components of the image. The specific process of wavelet transform is as follows: each row of the feature map is low-pass filtered and high-pass filtered, then downsampled to obtain the low-frequency and high-frequency components in the horizontal direction; subsequently, each column of the low-frequency and high-frequency components is low-pass filtered, high-pass filtered, and downsampled, finally obtaining four sub-bands: LL, LH, HL, and HH.

[0034] Based on the characteristics of the three high-frequency subbands, a high-frequency processing module is designed, such as... Figure 2As shown, a spatial-channel dual attention mechanism is adopted. The design of the channel interaction branch references the squeeze-excitation idea of ​​SENet, adaptively modeling the channel dependencies of high-frequency subband feature maps. First, global average pooling and global max pooling are performed on the input feature maps respectively. Global average pooling is responsible for global statistical information, while global max pooling is responsible for local salient features. The two are complementary. At the same time, more accurate channel weight features are obtained to capture the overall distribution statistics and the most salient response features in the channel dimension, resulting in two global channel descriptors. Then, the two channel descriptors are input into a shared MLP bottleneck structure for modeling. This MLP consists of two 1×1 convolutional layers: first, the channel compression ratio is set to 16 to promote cross-channel information interaction, and then the channel dimension is restored to generate channel responses, with the ReLU activation function used in between. Finally, the channel attention weights are obtained through the Sigmoid function.

[0035] , Where F is the shared MLP, GAP(X) is global average pooling, GMP(X) is global max pooling, and channel dimension weighting is applied to the high-frequency subband feature map to highlight the effective channels that are sensitive to high-frequency details and suppress noise interference channels.

[0036] The spatial interaction branch aims to model the spatial saliency of high-frequency subband feature maps at the pixel level. This branch uses 1×1 convolutions to aggregate information along the channel dimension, mapping multi-channel high-frequency features to a single-channel spatial response map, thus showing the importance of modeling different spatial locations. Spatial attention weights are then generated using the Sigmoid function.

[0037] , The input feature map is weighted pixel-wise to enhance high-frequency salient regions such as textures and edges, while suppressing spatially concentrated noise areas. After completing attention modeling in both channel and spatial dimensions, the channel-weighted feature map and the spatial-weighted feature map are added element-wise to achieve parallel and collaborative fusion of the two attention mechanisms. The fused features are further integrated and regularized using 3×3 convolution and Group Normalization to obtain enhanced high-frequency sub-band feature representations, thus maintaining both structural continuity and detail enhancement. Unlike sequential attention mechanisms, this parallel fusion strategy avoids mutual interference between spatial and channel attention, enabling the network to adaptively adjust spatial saliency and channel response intensity while maintaining the integrity of the high-frequency structure.

[0038] To address the global characteristics of the low-frequency subband, a low-frequency processing module is designed, such as... Figure 3As shown, a hybrid CNN-Mamba architecture is adopted to balance local details and global contextual information. First, a 3×3 convolutional layer is used to extract local features from the low-frequency subband feature map, followed by Group Normalization for normalization to enhance the stability of feature representation and preserve the image's basic structure and smooth region information. Then, the obtained two-dimensional low-frequency feature map is rearranged into a one-dimensional sequence, resulting in a dimension-adapted Mamba module modeling the sequence input. To fully model the global spatial dependencies in the low-frequency features, a bidirectional Mamba structure is introduced to process the sequence features. Specifically, the original sequence is input into the Mamba module for forward scanning, and simultaneously, the sequence is flipped spatially and input into the same Mamba module for reverse scanning, with the reverse output restored to the original order. Finally, the outputs of the forward and reverse scans are added element-wise to achieve bidirectional contextual information fusion. Through this bidirectional scanning mechanism, each sequence position can aggregate contextual information from the global scope, effectively modeling the long-range spatial dependencies in the low-frequency subband. Finally, the one-dimensional sequence features enhanced by the Mamba module are reshaped back into a two-dimensional spatial structure to obtain the enhanced low-frequency subband feature map. The core computation of the Mamba module is based on a structured state-space model, and its discretized state update process can be represented as:

[0039] in Indicates a hidden state. Given the input sequence features, A, B, C, and D are learnable parameter matrices. In the specific Mamba implementation, the dimensions and initialization of A, B, C, and D are determined by the hidden dimension and the channel dimension, and are learned through parameterization. Matrix A has an N×N dimension and is initialized with small negative values ​​close to those of a stable system; matrix B has an N×d dimension and is initialized using Xavier or Kaiming; matrix C has a d×N dimension and is initialized using Xavier or Kaiming; matrix D has a d×d dimension and is initialized with a constant 1 or a learnable parameter. This structure can effectively model global dependencies in long sequences while maintaining linear computational complexity.

[0040] After obtaining the enhanced low-frequency and high-frequency subband features, they are input into the inverse discrete wavelet transform module, such as... Figure 4 As shown, the frequency domain features are reconstructed back to the spatial domain through filtering and upsampling operations, thus obtaining the feature map enhanced by the WMH module, which provides a high-quality feature representation for subsequent reconstruction or enhancement tasks.

[0041] Discrete wavelet transform and inverse transform can be expressed as: , , Where LL represents the low-frequency subband, LH, HL, and HH represent high-frequency subbands in different directions, and LL', LH', HL', and HH' represent the enhanced subband characteristics, respectively.

[0042] The bidirectional Mamba global modeling process can be represented as follows: the low-frequency subband feature map is flattened to obtain a one-dimensional sequence. The bidirectional Mamba processing procedure is defined as follows: , , , Where M represents a Mamba module that shares the same parameters, and Flip represents a reverse operation along the sequence dimension. The fused bidirectional contextual feature sequence For forward scan output, This is the output for backward scanning.

[0043] For model training, commonly used low-light-normal-light paired datasets such as LOLv1, LOLv2-real, and LOLv2-syn are used, containing scene images with different indoor and outdoor lighting intensities, for model training and validation. The loss function is designed as a composite loss function, combining L1 loss, SSIM loss, and total variation regularization, as shown in the following formula:

[0044] , in, Used to optimize pixel-level precision Used to maintain image structural similarity This is used to suppress image noise and artifacts, and improve spatial smoothness. The weight parameters of 0.95 and 0.05 were determined based on preliminary experimental experience and can also be adjusted according to different augmentation needs of different dataset types, aiming to ensure that the reconstruction term has a higher weight and the TV term has a lower weight. The training parameters are set using the Adam optimizer, with momentum set to 0.9, batch size of 64, initial learning rate of 0.0013, decaying to 1e-7 using a cosine annealing strategy, 2500 epochs of training, and a CUDA-enabled GPU device.

[0045] Through the above technical solution, the present invention achieves high-quality enhancement of low-light images, effectively maintaining detail, texture and color naturalness while improving image brightness, and has good cross-scene adaptability.

[0046] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A low-light image enhancement method based on frequency domain decoupling and state space model, characterized in that, Includes the following steps: Acquire a low-light input image and preprocess it to obtain a feature map; The feature map is input into a pre-trained augmentation network, which includes multiple encoding units, a latent layer and multiple decoding units connected in sequence, with each encoding unit corresponding to a decoding unit. The feature map is first input into the first layer coding unit, and then enhanced by the WMH module to obtain an enhanced feature map. The enhanced feature map is then input into the next layer coding unit after the feature map resolution is gradually reduced and the number of channels is increased by downsampling between adjacent coding units. The above operation is repeated until the deep feature map output by the last coding unit is input into the latent layer after downsampling. The latent layer performs frequency domain adjustment on the input deep feature map to obtain an optimized feature map; Before being input into the first decoding unit, the optimized feature map is upsampled to restore the resolution. The resulting upsampled feature map is then weighted and fused with the enhanced feature map output by the corresponding encoding unit through skip connections. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The above upsampling operation is repeated until the feature map output by the last decoding unit is convolved to obtain the enhanced image. Specifically, the following sub-steps are performed in the WMH module of each encoding unit and each decoding unit: The input feature map is decomposed into low-frequency sub-bands and high-frequency sub-bands using discrete wavelet transform; For the low-frequency sub-band, a hybrid CNN-state space model module is used to extract local features and model the global context to obtain enhanced low-frequency features; A parallel spatial-channel dual attention module is used to enhance texture and suppress noise in the high-frequency subband to obtain enhanced high-frequency features. The enhanced low-frequency and high-frequency features are reconstructed into an enhanced feature map through inverse discrete wavelet transform and then output.

2. The low-light image enhancement method based on frequency domain decoupling and state space model of claim 1, wherein, The WMH module includes: Discrete wavelet transform unit is used to decompose the input feature map into a low-frequency subband and three high-frequency subbands in three directions; The high-frequency processing submodule is used to perform spatial-channel dual attention enhancement on the three high-frequency subbands to obtain enhanced high-frequency features. The low-frequency processing submodule is used to perform CNN local feature extraction and bidirectional state-space model global modeling on the low-frequency sub-band to obtain enhanced low-frequency features. The inverse discrete wavelet transform unit is used to reconstruct the enhanced low-frequency subband and high-frequency subband into enhanced feature maps and output them.

3. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 2, characterized in that, The high-frequency processing submodule employs a parallel spatial attention and channel attention mechanism: The channel attention branch extracts channel descriptors through global average pooling and global max pooling, and generates channel weights through shared MLP; The spatial attention branch generates spatial weights by aggregating channel information through convolution; The enhanced high-frequency features are obtained by adding the channel-weighted features and spatial-weighted features element by element, followed by convolution and normalization.

4. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 2, characterized in that, The low-frequency processing submodule includes: The CNN local feature extraction layer is used to extract local structural information of low-frequency subbands and output a two-dimensional low-frequency feature map. A serialization unit is used to flatten the two-dimensional low-frequency feature map into a one-dimensional sequence. A bidirectional state-space model unit is used to perform forward and backward scanning on the one-dimensional sequence, fuse bidirectional context information, and output an enhanced one-dimensional sequence. The deserialization unit is used to restore the enhanced one-dimensional sequence to a two-dimensional feature map. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 4 is characterized in that the bidirectional state-space model unit adopts a Mamba structure, and its state update process is expressed as follows: ， ， where x t is the input sequence, h t is the hidden state, y t is the output, and A, B, C, D are learnable parameters.

5. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 1, characterized in that, The number of the plurality of coding units is the same as the number of the plurality of decoding units. The skip connection of the j-th layer decoding unit receives the feature map output by the j-th layer coding unit, where j=1,2,...,N, N is the total number of coding units, and the first layer consists of the first input coding unit and the last output decoding unit.

6. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 1, characterized in that, The latent layer of the enhanced network is equipped with a frequency domain processing module, which is used to adjust the frequency domain of the deep feature map output by the encoder and optimize the expression of texture information.

7. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 1, characterized in that, The discrete wavelet transform and inverse discrete wavelet transform satisfy the following relationship: ， ， Where LL represents the low-frequency subband, LH, HL, and HH represent high-frequency subbands in different directions, and LL', LH', HL', and HH' represent the enhanced subband characteristics, respectively. For the input feature map, The reconstructed feature map.

8. The low-light image enhancement method based on frequency domain decoupling and state-space model as described in claim 1, characterized in that, The augmented network is trained using a composite loss function, which is expressed as follows: ， in, Used to optimize pixel-level precision Used to maintain image structural similarity Used to suppress image noise and artifacts, 0.95 and 0.05 are weighting coefficients.

9. A low-light image enhancement system based on frequency domain decoupling and a state-space model, characterized in that, include: The image acquisition module is used to acquire low-light input images and preprocess the images to obtain feature maps; An enhancement processing module is used to load pre-trained enhancement networks; The enhancement processing module includes: Multiple coding processing units are embedded with WMH modules. Each coding processing unit is used to receive feature maps after downsampling to reduce resolution and increase the number of channels. The feature maps are then enhanced by the WMH module to obtain enhanced feature maps, which are then downsampled and input into the next coding unit. The deep feature maps output by the last coding unit are downsampled and input into the latent layer. The latent layer processing unit is used to perform frequency domain adjustment on the input deep feature map to obtain an optimized feature map, which is then upsampled and input to the first decoding unit. Multiple decoding processing units are embedded with WMH modules. Each decoding unit is used to receive the upsampled feature map obtained by upsampling to restore resolution, and then perform skip connection weighted fusion of the upsampled feature map and the enhanced feature map output by the corresponding encoding unit. After fusion, it is enhanced by the WMH module and then upsampled again before being input into the next decoding unit. The feature map output by the last decoding unit is convolved to obtain the enhanced image. The WMH modules all include: Discrete wavelet transform unit is used to decompose the input feature map into low-frequency subband and high-frequency subband; The low-frequency processing submodule is used to perform local feature extraction and global context modeling on the low-frequency subband using a hybrid CNN-state space model module to obtain enhanced low-frequency features. The high-frequency processing submodule is used to perform texture enhancement and noise suppression on the high-frequency subband using a parallel spatial-channel dual attention module to obtain enhanced high-frequency features. The inverse discrete wavelet transform unit is used to reconstruct the enhanced low-frequency features and high-frequency features into an enhanced feature map and output it. The output module is used to output the enhanced image.

Images