An improved dual-frequency-domain feature aggregation adaptive low-illumination image enhancement method

Abstract

The application discloses a kind of based on improved dual-frequency domain feature aggregation adaptive low-illumination image enhancement method, this method can effectively solve the problems such as large low-illumination image noise, texture loss and color distortion.The method comprises the following steps: 1) construct low-illumination image dataset and carry out pretreatment;2) construct dual-flow enhancement network (named dual-frequency domain adaptive transformer network, DFAT-Net) including adaptive spectral illumination estimation branch and wavelet domain global perception branch;3) train adaptive spectral illumination estimation branch, generate gate weight correction phase spectrum using the stability of amplitude spectrum, utilize the strong coupling relationship between amplitude and phase to distinguish and suppress random noise in frequency domain, and restore illumination structure without loss;4) train wavelet domain global perception branch, separate high-frequency texture using wavelet transform, and capture non-local similarity through Transformer, distinguish and restore weak texture details submerged by noise from high-frequency subband;5) train dual-domain feature dynamic aggregation module, adaptively fuse frequency domain and time-frequency domain features;6) end-to-end optimization is carried out to network in combination with hybrid loss function.The application significantly improves the quality and visual naturalness of low-illumination image by improving frequency domain interaction and introducing global perception ability, effectively solves the problems such as noise residue, detail erasure and color shift in low-illumination image enhancement.

Description

Technical Field

[0002] This invention relates to the fields of computer vision and image processing technology, and specifically to an adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation that combines Fourier transform and wavelet transform. Background Technology

[0004] In scenarios such as nighttime surveillance and autonomous driving, the acquired images often suffer from low brightness, poor contrast, and high noise levels. Existing low-light image enhancement methods are mainly divided into traditional methods and deep learning-based methods. Traditional methods, such as Retinex theory, can improve brightness but are prone to color distortion and noise amplification. While deep learning-based methods (such as CNNs) have improved performance, most currently only process in the spatial domain, neglecting the importance of frequency domain information for denoising and decoupling illumination components. Although low-light image enhancement methods based on joint processing of dual frequency domains (Fourier domain and wavelet domain) (such as DF-DFANet) have emerged in recent years, attempting to resolve the contradiction between illumination enhancement and noise suppression through frequency domain separation, existing technologies still face the following significant technical bottlenecks in practical applications: In the Fourier transform, the amplitude spectrum mainly reflects the energy information of the image, such as brightness and contrast, while the phase spectrum carries structural information such as edges and contours. Existing frequency domain enhancement methods typically employ a local enhancement strategy, i.e., designing two independent branches to process the amplitude and phase spectra separately, or simply using complex convolution to process them in parallel. However, this approach ignores the strong coupling between amplitude and phase. In low-light enhancement, significant brightness increases (amplitude spectrum changes) often require corresponding structural information (phase spectrum) for adaptation. Current technologies lack an effective cross-component interaction mechanism, failing to dynamically guide phase spectrum repair using amplitude spectrum trends or correct over-enhancement of the amplitude spectrum using phase spectrum structural constraints. This approach easily leads to mismatches between illumination distribution and object structure during reconstruction, severely impacting the image's visual naturalness. Low-light images are often accompanied by extremely low signal-to-noise ratios and severe noise interference, easily obscuring realistic texture details with high-frequency noise. Local convolution operations alone cannot distinguish between weak texture signals and random noise, causing the network to tend to treat high-frequency information as noise for smoothing, resulting in over-smoothing of image details. Furthermore, images often contain a large amount of non-local similarity. Existing CNN architectures struggle to establish long-distance dependencies between pixels and cannot utilize global contextual information to assist in the reconstruction of local textures. Ultimately, this results in enhanced images that, while brighter, have blurred details and lack realism. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing low-light image enhancement techniques and provide an adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation. This method effectively improves the model's denoising and detail recovery capabilities by improving the frequency domain processing module and introducing the Transformer mechanism.

[0007] Compared with the prior art, the present invention has the following beneficial effects and working principle:

[0008] 1. Regarding how to solve the problem of excessive noise:

[0009] This invention achieves denoising through an amplitude-phase interaction mechanism in the adaptive spectral illumination estimation branch. In low-light images, noise typically manifests as high-frequency random signals, which severely disrupt the phase spectrum in the frequency domain, leading to structural information corruption. The amplitude spectrum, however, primarily carries illumination and contrast energy and is relatively stable. This invention innovatively utilizes relatively clean amplitude spectrum features to generate gating weights, performing element-wise weighted correction on the noise-affected phase spectrum. The principle is to leverage the significant energy distribution in the amplitude spectrum to enhance the corresponding phase information while suppressing phase fluctuations in regions with low amplitude energy. This cross-component interaction is essentially a frequency-domain soft-thresholding denoising method, suppressing color noise and random noise in dark areas at their source, rather than relying solely on subsequent smoothing convolutions.

[0010] 2. Regarding how to solve the problem of texture loss.

[0011] This invention addresses texture loss and over-smoothing issues through a wavelet domain global perception branch. Traditional convolutional neural networks tend to erase subtle texture details as noise during denoising. The principle of this invention is as follows: First, the image is decomposed using Discrete Wavelet Transform (DWT), explicitly separating low-frequency and high-frequency sub-bands. Texture details are primarily concentrated in the high-frequency sub-band. A multi-head self-attention mechanism from the Transformer is introduced for the high-frequency sub-band. Because image textures exhibit non-local similarity (i.e., texture in one part of the image may be similar to texture in another distant part), the self-attention mechanism captures this long-distance dependency, utilizing clear texture features from other areas of the image to assist in reconstructing the texture of the currently blurred region. This allows the network to remove random noise while preserving and enhancing texture details with repetitive patterns.

[0012] 3. Regarding how to solve the problem of color distortion

[0013] This invention addresses color distortion through a hybrid loss function combining dynamic aggregation of dual-domain features and physical constraints. The principle is as follows: Existing enhancement methods often process RGB channels independently or excessively linearly amplify brightness, leading to color shift. This invention first adjusts the illumination in the frequency domain to ensure overall brightness consistency; secondly, in the feature aggregation module, it uses a spatial attention map to dynamically balance the fusion ratio of illumination features and detail features, avoiding color whitening caused by local overexposure. During the optimization stage, frequency domain consistency loss and perceptual loss are introduced to force the enhanced image to maintain consistency with the normally illuminated image in terms of frequency distribution and high-level semantic features. This improves brightness while correcting color deviations caused by insufficient sensor sensitivity, restoring true visual colors.

[0014] This invention includes the following main steps:

[0015] 1. A low-light semantic segmentation method based on AdaptGAN, characterized by the following steps:

[0016] (1) Dataset construction: Based on the publicly available large-scale PASCAL VOC 2012 dataset, the corresponding synthetic low-light dataset PASCAL-Low was obtained by simulating the imaging characteristics of real nighttime low-light environments (including nonlinear illumination attenuation and sensor noise injection). The original PASCAL VOC dataset was used as the normal illumination domain, and the PASCAL-Low dataset was used as the low-light domain. A total of 17,125 pairs of images were divided into training and testing sets in a 7:3 ratio. (2) A dual-stream enhancement network model was constructed. This model includes a parallel adaptive spectral illumination estimation branch and a wavelet domain global perception branch, as well as a dual-domain feature dynamic aggregation module. (3) The adaptive spectral illumination estimation branch was trained, including the following sub-steps: (3.1) The input image was subjected to fast Fourier transform to convert the image from the spatial domain to the frequency domain and separate the amplitude spectrum features and phase spectrum features. (3.2) The amplitude-phase interaction unit was trained. The separated amplitude spectrum features and phase spectrum features are fed into the interaction unit. The global energy information of the amplitude spectrum is used to gate the structural information of the phase spectrum. This step is based on the frequency domain imaging principle, that is, the amplitude spectrum mainly carries the brightness and contrast information of the image, while the phase spectrum mainly carries the edge structure information. By establishing an explicit guidance mechanism of amplitude to phase, high-frequency random noise in the phase spectrum of dark areas can be adaptively suppressed according to the intensity of illumination, thereby solving the common color distortion and edge ringing effect in low illumination enhancement from the source. (3.3) The frequency domain features after interaction processing are processed through stacked complex convolutional layers for feature extraction and fine adjustment. (3.4) The processed frequency domain features are subjected to fast inverse Fourier transform to restore the features to the spatial domain, resulting in an illumination feature map containing global illumination information. (4) The wavelet domain global perception branch is trained, including the following sub-steps: (4.1) The input image is subjected to discrete wavelet transform to decompose it into a low-frequency sub-band and three high-frequency sub-bands in the horizontal, vertical and diagonal directions. (4.2) The four sub-bands are concatenated along the channel dimension as input features. (4.3) The wavelet transform module is trained. The concatenated sub-band features are fed into the wavelet transform module, which uses a multi-head self-attention mechanism to capture long-distance dependencies and non-local texture features in the image, solving the problem of limited receptive field in traditional convolution. Due to the extremely low signal-to-noise ratio of low-light images, it is difficult to distinguish weak textures from noise using only local convolution. This invention utilizes the global modeling capability of the Transformer to mine the non-local similarity of the image in the high-frequency sub-bands after wavelet decomposition, that is, to use clear and repetitive texture features in other regions of the image to assist in the reconstruction of details in the current region, effectively avoiding the problem of excessive texture smoothing caused by traditional denoising methods.(4.4) Perform discrete wavelet inverse transform on the features output by the wavelet transform module to restore the detail feature map with the same resolution as the input image. (5) Train the dual-domain feature dynamic aggregation module, the steps are as follows: (5.1) First, concatenate the illumination feature map obtained in (3.4) and the detail feature map obtained in (4.4) to form a hybrid feature tensor. (5.2) Use convolutional layers to perform spatial attention modeling on the hybrid features, and generate two complementary spatial weight maps through the Sigmoid activation function, corresponding to the frequency domain illumination component and the wavelet domain detail component, respectively. (5.3) Multiply and add the generated weight map element-wise with the corresponding input feature map to achieve pixel-level adaptive fusion. (5.4) After fusion, the features are processed by convolutional layers to adjust the channels and output the preliminary enhanced image. This module breaks the traditional simple linear superposition method and can dynamically adjust the fusion ratio according to the texture richness of the image region. (6) Training the global feature refinement module, the steps are as follows: (6.1) The preliminary enhanced image output in step (5.4) is used as input. First, it passes through a shallow convolutional layer for feature mapping, mapping the image space into a high-dimensional feature space. (6.2) The feature map passes through continuous downsampling encoding blocks, extracting multi-scale deep semantic features while reducing the resolution, and expanding the receptive field to capture the global illumination distribution. (6.3) Dense connection blocks are used in the bottleneck layer with the lowest resolution to strengthen feature transfer, improve feature reuse rate and prevent gradient vanishing during training. (6.4) Spatial resolution is gradually restored through bilinear interpolation upsampling, and skip connections are used to fuse the shallow texture features in the encoding stage with the deep features in the decoding stage to compensate for the information loss caused by downsampling. (6.5) The decoded feature map is mapped back to the image pixel space through the final convolutional layer and nonlinear activation function, outputting a high-quality enhanced image with a natural visual feel. (7) Calculate the hybrid loss between the enhanced image and the normal illumination label image, including L1 reconstruction loss, perceptual loss and frequency domain consistency loss. Use the backpropagation algorithm and Adam optimizer to update the network parameters end-to-end by minimizing the hybrid loss function.

[0017] 2. The adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation according to claim 1 is characterized in that the interactive processing of amplitude spectrum and phase spectrum in step (3.2) is as follows: First, global average pooling is performed on the input amplitude spectrum features to compress the two-dimensional feature map into a one-dimensional global description vector; second, the vector is fed into two fully connected layers (FC) to model the inter-channel dependency relationship, where the first fully connected layer is used for dimensionality reduction and the second fully connected layer is used for dimensionality recovery; then, the Sigmoid activation function is used to map the output to weight coefficients between 0 and 1; finally, the generated weight coefficients are multiplied element-wise with the input phase spectrum features and the residuals are added to the original phase spectrum features to obtain the corrected phase spectrum features guided by amplitude information. This technical solution addresses the problem of edge artifacts, texture loss, and color distortion in low-light image enhancement, where existing technologies neglect the coupling relationship between amplitude and phase in the frequency domain and the limited receptive field of convolution in the wavelet domain. In this technical solution, a large-scale synthetic low-light dataset is first constructed to ensure the model's generalization ability. Then, through an adaptive spectral illumination estimation branch, an amplitude-phase interaction mechanism is introduced in the frequency domain, using amplitude information to dynamically guide the recovery of phase information, thereby ensuring the consistency between illumination enhancement and structural reconstruction. Simultaneously, a wavelet domain global perception branch is used, employing a Transformer structure to replace traditional convolution, capturing global contextual information in non-overlapping wavelet high-frequency subbands to achieve high-fidelity restoration of weak textures. Next, a dual-domain feature dynamic aggregation module adaptively fuses illumination and detail features. Finally, a global feature refinement module eliminates potential block artifacts from feature fusion and optimizes the results using a hybrid loss function. This method significantly improves the brightness and contrast of low-light images, enabling the enhanced image to effectively suppress noise while retaining richer texture details, resulting in a more natural and realistic visual effect. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the DFAT-Net model in an embodiment of the present invention;

[0020] Figure 2 This is a structural diagram of the adaptive spectral illumination estimation branch in an embodiment of the present invention;

[0021] Figure 3 This is a detailed internal structure diagram of the amplitude-phase interaction unit in an embodiment of the present invention;

[0022] Figure 4 This is a structural diagram of the wavelet domain global sensing branch in an embodiment of the present invention;

[0023] Figure 5This is a detailed internal structure diagram of the wavelet transform module in an embodiment of the present invention;

[0024] Figure 6 This is a structural diagram of the dual-domain feature dynamic aggregation module in an embodiment of the present invention;

[0025] Figure 7 is a structural diagram of the global feature refinement module in an embodiment of the present invention; Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the following description, in conjunction with the accompanying drawings and embodiments, further illustrates the invention. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0028] refer to Figure 1 An adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation includes the following steps:

[0029] (1) Dataset construction: Based on the publicly available large-scale PASCAL VOC 2012 dataset, the imaging characteristics (including gamma correction coefficients) of real nighttime low-light environments were simulated. Nonlinear illumination attenuation and variance (Injecting Poisson-Gaussian noise) to obtain the corresponding synthetic low-light dataset PASCAL-Low. In this invention, the original PASCAL VOC dataset is used as the normal illumination domain and the PASCAL-Low dataset is used as the low-light domain. A total of 17,125 pairs of images are divided into training and testing sets in a 7:3 ratio. (2) Construct and train the Dual-Frequency Adaptive Transformer Network (DFAT-Net). The network adopts a parallel dual-stream architecture and is mainly composed of an adaptive spectral illumination estimation branch and a wavelet domain global perception branch. The overall network structure of DFAT-Net is shown in Figure 1. (3) Train the adaptive spectral illumination estimation branch. The structure of the adaptive spectral illumination estimation branch is as follows. Figure 2 As shown, the processing procedure includes the following sub-steps: (3.1) The input image with a size of 416×416×3 is fed into a preprocessing convolutional block. This convolutional block contains a convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, increasing the number of input channels from 3 to 64, followed by a LeakyReLU activation layer with a parameter of 0.2. Then, the extracted spatial features are processed... Perform a Fast Fourier Transform (FFT), the formula is as follows: The obtained complex feature map is separated into real part R and imaginary part I, and the amplitude spectrum is further calculated. and phase spectrum The magnitude spectrum A and the phase spectrum P are both 416×416×64. (3.2) The magnitude spectrum A and the phase spectrum P are fed into the magnitude-phase interaction unit for processing. The detailed structure of the magnitude-phase interaction unit is shown in Figure 3. First, the magnitude spectrum A is input into a global average pooling layer to obtain a global description vector with 64 channels; then the vector is fed into the first fully connected layer with 64 input nodes and 16 output nodes, and the activation function is ReLU; then it is fed into the second fully connected layer with 16 input nodes and 64 output nodes, and the activation function is Sigmoid to obtain the gated weight vector; finally, the weight vector is multiplied element-wise with the original phase spectrum P and added to the original phase spectrum P. The calculation formula is as follows: The corrected phase spectrum P′ is obtained. The specific mechanism for denoising here is that noise in low-light images manifests as chaotic phase jumps in the phase spectrum. The weight vector generated by the Sigmoid function... It's actually a noise suppression mask. It's applied at image edges and in areas with large amplitude values ​​(primarily around main objects). It tends towards higher values, preserving and enhancing the phase structure; in flat dark areas (small amplitude values, noise dominates). The phase response tends to be low, thus attenuating the phase response in that region. This amplitude-based phase correction effectively filters out noise components in the frequency domain, laying a low-noise foundation for subsequent image recovery. (3.3) The amplitude spectrum A and the corrected phase spectrum P′ are concatenated and fed into three stacked complex convolutional residual blocks. This avoids the inability of traditional convolutional neural networks to directly process complex data. Assume the input frequency domain features are Each residual block contains two complex convolutional layers with convolutional kernels of type 1. All of them are 3×3 in size, with 64 input and output channels, a stride of 1, and padding of 1. In complex convolutional layers, the actual operations performed follow the distributive law of complex multiplication: Where * denotes the standard convolution operation. Through this cross-convolution of the real and imaginary parts, the network can simultaneously learn amplitude adjustment and phase shift in the frequency domain, thereby improving brightness while maintaining the sharpness of image edges and avoiding the "ringing effect" common in traditional methods. The convolution is followed by the ComplexReLU activation function. (3.4) The real and imaginary features after processing through 5 complex residual blocks are recombined into complex form and then subjected to inverse fast Fourier transform (IFFT) to restore the illumination feature map in the spatial domain. The number of output channels is 64. (4) Train the wavelet domain global perception branch, which is used to capture global texture details of the image. The structure of the wavelet domain global perception branch is as follows: Figure 4 As shown, the processing procedure includes the following sub-steps:

[0030] (4.1) Perform Haar wavelet transform on the input image. To obtain a larger receptive field, this embodiment adopts a two-stage wavelet decomposition strategy (increasing depth). First, perform the first-stage decomposition to obtain... Only for Perform a second-level decomposition to obtain Then, three primary sub-bands containing the richest high-frequency information were selected. As input, concatenate them along the channel dimension. The number of input channels is... (4.2) A convolutional layer with a kernel size of 3×3 is used to map the concatenated subband features to embedded features with C=96 channels. To preserve spatial location information, a learnable absolute positional encoding is introduced and added element-wise to the embedded features. (4.3) The feature map is fed into four stacked wavelet transform modules. The detailed structure of the wavelet transform modules is shown in Figure 5. First, the input features are processed... Perform layer normalization, the calculation formula is as follows: A query Q, key K, and value V matrix is ​​generated using 1×1 convolutions. The feature map channels are divided into 8 heads. Within each head, an attention score is calculated. In this step, an attention score is calculated. In practice, it achieves the key functions of texture restoration and noise suppression. The principle is as follows: noise in low-light images typically exhibits random distribution characteristics, with extremely low long-distance correlation between pixels, resulting in a small inner product score between the query and the key. In contrast, real texture details possess significant non-local self-similarity, enabling the network to find highly correlated feature matches globally, even under low-light interference, thus generating a high attention score. By weighting and aggregating this score with the value matrix V, the network adaptively enhances highly correlated texture feature responses while utilizing the non-linear properties of Softmax to suppress low-scoring random noise components. This effectively distinguishes texture from noise during feature extraction, solving the problem of texture detail loss caused by traditional convolution operations. Finally, the outputs of the eight heads are concatenated and fused through a linear layer. The output of MSA is then combined with... Add them together to obtain intermediate features. . The signal passes through LayerNorm again. Then it enters the FFN module, which contains two convolutional layers: the first layer expands the number of channels from 96 to 384, followed by the GELU activation function; the second layer restores the number of channels from 384 to 96. The output of FFN is added to Xmid to obtain the output of the current layer. (4.4) The features processed by 6 Transformer layers are adjusted back to 9 channels by a 3×3 convolutional layer, then divided into three sub-band components, and reconstructed into spatial domain detail feature maps using Inverse Discrete Wavelet Transform (IDWT). (5) Train the dual-domain feature dynamic aggregation module, which is used to fuse illumination and detail features. The structure of the dual-domain feature dynamic aggregation module is as follows: Figure 6 As shown, the specific steps are as follows: (5.1) The illumination feature map obtained in step (3.4) and the detail feature map obtained in step (4.4) are concatenated along the channel dimension to obtain a mixed feature map with 128 channels. (5.2) The mixed feature map is fed into a convolutional layer with a kernel size of 1×1 and an output channel of 2. Two spatial attention weight maps are generated by the Sigmoid function, corresponding to the illumination branch and the detail branch, respectively. (5.3) The two weight maps are multiplied element-wise with the corresponding input feature maps, and then the results are added together for fusion. (5.4) The fused features are passed through a convolutional layer with a kernel size of 3×3 and an output channel of 3 to obtain the final enhanced image. (6) Training the Global Feature Refinement Module. The structure of the Global Feature Refinement Module is shown in Figure 7. It includes the following sub-steps: (6.1) Input the enhanced image from step (5.4) into a convolutional layer with a kernel of 3×3, outputting 64 channels and using LeakyReLU as the activation function. (6.2) Pass the feature map through two consecutive downsampling blocks. Each block contains a 3×3 convolutional layer with a stride of 2 (for halving the resolution) and a residual block (for feature extraction). After two downsampling operations, the feature map size becomes 1 / 4 of its original size, and the number of channels increases to 128 and 256 respectively. (6.3) At the lowest resolution, feature propagation is enhanced by three dense blocks. Each dense block contains four convolutional layers with dense connections between them and a growth rate of 32.

[0031] (6.4) The feature map is processed through two consecutive upsampling blocks. Each block contains a bilinear interpolation layer to double the resolution, followed by a 1×1 convolutional layer to halve the number of channels, and then jump-connects and concatenates the feature map at the corresponding scale in the encoding stage. (6.5) The size of the decoded feature map is restored to 416×416, and finally it is processed by a convolutional layer with a kernel of 3×3 and 3 output channels, and then mapped to the image pixel space by the Tanh activation function to obtain the final high-quality enhanced image. In this process, color distortion is effectively solved because: the frequency domain processing in step (3) ensures that the enhancement of the luminance component will not destroy the topological structure of the image, thus avoiding structural color shift; at the same time, the refinement network in step (6) learns the mapping relationship from low-light hue to normal-light hue under the joint constraints of L1 loss and perceptual loss, and performs nonlinear correction on the RGB values ​​of the pixels, eliminating the color shift caused by simple linear amplification. (7) Calculate the loss function and optimize the network: (7.1) Calculate the L1 reconstruction loss of the enhanced image and the normal illumination label image, as shown in formula (1):

[0032] (1) where To enhance the image, For the label image, N is the total number of pixels.

[0033] (7.2) Calculate the perceptual loss, as shown in formula (2):

[0034] (2)

[0035] in Let represent the feature map of the j-th layer of the VGG-19 network. (7.3) Calculate the frequency domain consistency loss, as shown in formula (3):

[0036] (3)

[0037] in This represents the Fourier transform operation. (7.4) The final total loss is... The Adam optimizer was used for training.

Claims

1. An adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation, Its features include the following method steps: (1) Constructing a dataset: using the publicly available LOL (Low-Light) dataset or the MIT-Adobe FiveK dataset. Normalize the image size and scale it uniformly to 416×416 pixels. Randomly divide the dataset into training and testing sets; (2) Construct and train a dual-stream enhancement network, which includes a parallel adaptive spectral illumination estimation branch and a wavelet domain global perception branch, as well as a dual-domain feature dynamic aggregation module; (3) Train the adaptive spectral illumination estimation branch, including the following sub-steps: (3.1) Perform a fast Fourier transform (FFT) on the input image to convert the spatial domain image into a frequency domain representation and separate the amplitude spectrum A and the phase spectrum P; (3.2) Input the amplitude spectrum A and phase spectrum P into the amplitude and phase interaction unit. First, use 1×1 convolution to map the amplitude spectrum and phase spectrum to a high-dimensional feature space, and set the number of channels to 64; (3.3) In the interaction unit, use a gating mechanism to calculate the amplitude-phase modulation weights. Pass the amplitude features through a global average pooling layer and a fully connected layer (FC), and then use the Sigmoid activation function to generate a weight vector. Multiply the weight vector element-wise with the phase features to realize the guidance of amplitude information for phase recovery; (3.4) The amplitude spectrum and phase spectrum after interactive processing are respectively processed by three stacked complex convolutional residual blocks. Each residual block contains two 3×3 complex convolutional layers with a stride of 1 and padding of 1, and a ReLU activation function in between; (3.5) The enhanced frequency domain features are restored to the spatial domain illumination feature map using the inverse fast Fourier transform (IFFT); (4) The wavelet global perception branch is trained, and the steps are as follows: (4.1) The input image is subjected to discrete wavelet transform (DWT) to decompose it into a low-frequency sub-band (LL) and three high-frequency sub-bands (LH, HL, HH); (4.2) The four sub-bands are concatenated in the channel dimension, and the result is sent as input features to the wavelet transformer block; (4.3) The wavelet transformer block first extracts features through a 3×3 convolutional layer, outputting 128 channels, and then inputs the feature map into the multi-head self-attention (Multi-Head) module. In the Self-Attention unit, global texture details are enhanced by calculating long-distance dependencies between pixels; (4.4) The output of the Self-Attention unit passes through a feedforward neural network (FFN), which contains two 1×1 convolutional layers and a GELU activation function. The first convolutional layer expands the number of channels to 4 times, and the second convolutional layer restores the number of channels; (4.5) The processed feature map is subjected to discrete wavelet inverse transform (IDWT) to restore the detail feature map of the same size as the input image; (5) The dual-domain feature dynamic aggregation module is trained. The illumination feature map obtained in step (3.5) and the detail feature map obtained in step (4.5) are concatenated in the channel dimension. The number of channels is reduced to 64 through a 1×1 convolutional layer. Then, a spatial attention map is generated through the Sigmoid function. The two features are weighted and fused. Finally, the final enhanced image is output through a 3×3 convolutional layer. (6) Train the global feature refinement module. The steps are as follows: (6.1) The preliminary enhanced feature map output by the dual-domain feature dynamic aggregation module is fed into the refinement network based on the encoder-decoder structure. (6.2) Multi-scale deep features are extracted through continuous downsampling operations, and dense connection blocks are used in the bottleneck layer to strengthen feature transmission. (6.3) Use skip connections to fuse shallow texture features with upsampled deep features, and output a high-quality enhanced image through the final convolutional layer; (7) Calculate the loss between the enhanced image and the normal illumination label image, and update the network weights using the backpropagation algorithm.

2. The adaptive low-light image enhancement method based on improved dual-frequency domain feature aggregation as described in claim 1, characterized in that... The loss function in step (6) The definition is as follows: in, for Reconstruction loss is used to ensure pixel-level consistency. This is a perceptual loss function based on the VGG-19 network, used to recover the detailed texture structure of visual perception. To calculate the frequency domain consistency loss, the distance between the amplitude spectrum and phase spectrum of the enhanced image and the label image in the frequency domain is calculated. , , These are the weighting coefficients for the loss term, set as follows: =1, =0.1, =0.1.

Images