Multi-stage information interactive weak light image enhancement method based on double-channel normalized flow

Abstract

The application discloses a multi-stage information interactive weak light image enhancement method based on a double-channel normalized flow, and specific steps are as follows: step 1, converting a weak light RGB image to a YUV color space to obtain a luminance channel image Y and color channel images U and V; step 2, constructing an overall network, inputting the luminance channel image Y and the color channel images U and V obtained into the overall network to obtain corrected luminance channel image Y'', color channel image U'' and color channel image V''; and step 3, converting the corrected luminance channel image Y'', the color channel image U'' and the color channel image V'' obtained into an RGB image to obtain an enhanced RGB image. The method improves the enhancement effect on a weak light image.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image enhancement methods, and particularly relates to a multi-stage information interactive weak light image enhancement method based on a double-channel standardized flow. BACKGROUND

[0002] Weak light image enhancement is a key technology in the field of computer vision and image processing, and is widely used in mobile phone photography, video monitoring, automatic driving, medical imaging and astronomical image observation and many other scenes. In the environment with insufficient natural or artificial lighting, the images collected by the imaging device usually have problems such as low brightness, color distortion, blurred details, and serious noise interference. Such images not only affect the subjective experience of users, but also significantly reduce the accuracy of downstream tasks such as target detection, recognition and segmentation. Therefore, it is of great theoretical value and practical significance to design an efficient and generalizable weak light image enhancement method.

[0003] Traditional weak light image enhancement methods mainly fall into three categories: image processing methods, physical modeling methods, and deep learning-based methods. The first category of methods is based on image processing enhancement methods, such as Histogram Equalization (HE), CLAHE, Gamma Correction, and Retinex theory. This type of method enhances the overall visual effect by enhancing the contrast of the image or adjusting the brightness channel of the image, and can achieve certain enhancement effect in simple scenes. The second category of methods attempts to construct an image degradation model, and realizes image reconstruction by inverting the imaging process. For example, the Retinex-based model decomposes the image into a reflection component and an illumination component, and enhances the illumination part to improve the overall brightness of the image. This type of method has certain interpretability for low-light images in theory. The third category of methods is based on convolutional neural networks (CNN) or generative adversarial networks (GAN) for weak light image enhancement. This type of method can extract statistical rules from large-scale images through end-to-end learning to improve image quality. Typical works include Zero-DCE series, EnlightenGAN, RetinexNet, KinD, etc. These models usually use U-Net, ResNet, or Transformer structure for image brightness restoration and detail enhancement. Some methods also consider denoising and color correction.

[0004] However, the above methods all have technical bottlenecks in color modeling: the U channel (blue-yellow component) and the V channel (red-green component) are often severely disturbed by noise in weak light environments, and their distribution deviates significantly from natural images. Existing methods generally ignore the correlation between the U and V channels and do not utilize their potential statistical dependencies, which leads to color distortion in the enhanced images. SUMMARY

[0005] The purpose of the present application is to provide a multi-stage information interactive low-light image enhancement method based on a two-channel normalized flow, solving the problem of color distortion of the enhanced image existing in the existing low-light image enhancement method.

[0006] The technical scheme adopted by the present application is a multi-stage information interactive low-light image enhancement method based on a two-channel normalized flow, and the specific steps are as follows:

[0007] Step 1, convert the low-light RGB image to the YUV color space to obtain the luminance channel image Y and the color channel image U and the color channel image V;

[0008] Step 2, construct an overall network, input the luminance channel image Y and the color channel image U and the color channel image V obtained in step 1 into the overall network to obtain the corrected luminance channel image Y'', the color channel image U'', and the color channel image V''.

[0009] Step 3, convert the corrected luminance channel image Y'', the color channel image U'', and the color channel image V'' obtained in step 2 into an RGB image to obtain an enhanced RGB image.

[0010] The present application is also characterized in that:

[0011] In step 2, the overall network is composed of a luminance perception enhancement network, a two-channel Normalizing Flow module, and a joint correction and reconstruction network.

[0012] The luminance channel image Y is input into the luminance perception enhancement network, and the enhanced luminance channel image Y' is output.

[0013] The color channel image U and the color channel image V are input into the two-channel Normalizing Flow module, and the enhanced color channel image U' and the enhanced color channel image V' are output.

[0014] The enhanced luminance channel image Y', the enhanced color channel image U', and the enhanced color channel image V' are input into the joint correction and reconstruction network, and the corrected luminance channel image Y'', the color channel image U'', and the color channel image V'' are output.

[0015] The luminance perception enhancement network is composed of a multi-scale feature extraction module, a multi-scale feature fusion module, a luminance perception attention module, and a transposed convolution recovery module.

[0016] The processing procedure of the brightness perception enhancement network is: the brightness channel image Y is input into the multi-scale feature extraction module and the brightness perception attention module respectively, the output of the multi-scale feature extraction module is input into the multi-scale feature fusion module, the multi-scale feature fusion module outputs a multi-scale feature map, the brightness perception attention module outputs a brightness perception weight map, the multi-scale feature map and the brightness perception weight map are multiplied element by element to obtain a weighted brightness feature map, and the weighted brightness feature map is input into the transposed convolution recovery module to input the enhanced brightness channel image Y';

[0017] The multi-scale feature extraction module is composed of at least 3 convolution extraction modules, each convolution extraction module is sequentially composed of a convolution layer, a normalization layer (Batch Normalization) and a ReLU activation function;

[0018] The multi-scale feature fusion module includes a spatial alignment stage and a channel alignment stage;

[0019] The brightness perception attention module is sequentially composed of a global average pooling and an SE module;

[0020] The transposed convolution recovery module is composed of at least 2 transposed convolution modules, each transposed convolution module is sequentially composed of a transposed convolution layer, a ReLU activation and a normalization layer.

[0021] The dual-channel Normalizing Flow module is composed of a U-channel Flow module, a V-channel Flow module and two multi-stage feature information interaction modules;

[0022] The U-channel Flow module is sequentially composed of Stage1, Stage2 and Stage3;

[0023] The V-channel Flow module is sequentially composed of Stage4, Stage5 and Stage6;

[0024] Stage1, Stage2, Stage3, Stage4, Stage5 and Stage6 are each composed of a learnable normalization layer, a reversible 1x1 convolution layer and an affine coupling layer;

[0025] The affine coupling layer divides the input into A and B, A remains unchanged, B passes through a lightweight network to output B', and A and B' are spliced to obtain the final output;

[0026] Each multi-stage feature information interaction module is sequentially composed of a feature splicing layer, a multi-head self-attention module and a residual connection;

[0027] The feature splicing layer is composed of at least one convolution layer;

[0028] The residual connection is sequentially composed of a convolution with a kernel of 1*1 and a skip connection.

[0029] The processing procedure of the dual-channel Normalizing Flow module is as follows: the color channel image U is input into Stage 1, the color channel image V is input into Stage 4, the output features of Stage 1 and the output features of Stage 4 are input into the first multi-stage feature information interaction module, the output of the first multi-stage feature information interaction module is input into Stage 2 and Stage 5 respectively, the output features of Stage 2 and the output features of Stage 5 are input into the second multi-stage feature information interaction module, the output of the second multi-stage feature information interaction module is input into Stage 3 and Stage 6 respectively, Stage 3 outputs the enhanced color channel image U', and Stage 6 outputs the enhanced color channel image V'.

[0030] The joint correction and reconstruction network is composed of an encoder, a decoder and a skip connection;

[0031] The enhanced luminance channel image Y', the enhanced color channel image U' and the enhanced color channel image V' are input into the joint correction and reconstruction network, and the corrected luminance channel image Y'', the corrected color channel image U'' and the corrected color channel image V'' are output;

[0032] The encoder is composed of three convolution modules, each of which is sequentially composed of a convolution layer, a normalization layer and a ReLU activation function;

[0033] The decoder is sequentially composed of three transpose convolution modules and a Sigmoid activation; each of the transpose convolution modules is sequentially composed of a transpose convolution layer and a ReLU activation function;

[0034] Each convolution module in the encoder is connected to the corresponding transpose convolution module in the decoder by a skip connection.

[0035] In step 2, the total loss function of the overall network is:

[0036] (2)

[0037] In formula (2), is a reconstruction loss; is a maximum likelihood loss; is a color consistency loss; is a joint perceptual loss;

[0038] wherein,

[0039] (3)

[0040] In formula (3), is a structural similarity index, used to measure the structural fidelity of an image; and is a loss weight coefficient; is a real luminance image;

[0041] (4)

[0042] In formula (4), is the probability density of the Flow module after shooting under the standard Gaussian distribution, is a negative log-likelihood;

[0043] (5)

[0044] In formula (5), and is a real color channel image; is the color distribution of the enhanced color channel image U', is the color distribution of the normal light image channel, is a KL divergence, which measures the distance between two probability distributions; , and is a loss weight coefficient;

[0045] (6)

[0046] In formula (6), is a YUV space conversion function, is an enhanced RGB image, is a normal light image, indicates a perceptual loss weight.

[0047] The beneficial effects of the present application are:

[0048] (1) The multi-stage information interactive weak light image enhancement method based on a dual-channel normalized flow of the present application can specifically enhance the luminance of weak light regions, suppress noise regions and maintain image structure by constructing a luminance perceptual enhancement network, and such design can significantly reduce noise amplification and detail loss caused by merely balancing the whole, thereby obtaining high-quality luminance enhancement results.

[0049] (2) The multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application can provide independent and deep reversible modeling paths for color channel images U and V through the construction of a double-channel Normalizing Flow module, learn the nonlinear mapping between them and the natural image distribution, and solve the interference and ambiguity problems that occur when a pure convolutional network jointly models color components, so as to improve the color restoration degree and signal-to-noise ratio;

[0050] (3) The multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application can explicitly model the semantic dependency relationship between color channel images U and V through the construction of a multi-stage feature information interaction module, realize high exchange and alignment of information, reduce color cast and blur caused by semantic mismatch between color channel images U and V, and significantly improve the overall color restoration degree and signal-to-noise ratio of the restored image;

[0051] (4) The multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application can finally correct and finely adjust the overall image through the construction of a joint correction and reconstruction network, make up for the overall semantic and structural information that cannot be considered during separation processing, and significantly reduce the blocking effect caused by splicing at different stages, so as to output a final image with balanced brightness, natural color and clear structure;

[0052] (5) The multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application sets a multi-loss function joint supervision mechanism, simultaneously focuses on brightness accuracy, distribution fitting, color preservation and overall perceptual quality during training, and improves the convergence and generalization of the network model training process. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The flowchart of the multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application;

[0054] Figure 2 The structural diagram of the brightness perceptual enhancement network in the multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application;

[0055] Figure 3 The structural diagram of the double-channel Normalizing Flow module in the multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application;

[0056] Figure 4 The structural diagram of the joint correction and reconstruction network in the multi-stage information interactive weak light image enhancement method based on the double-channel normalizing flow of the application;

[0057] Figure 5 This is a visualization comparison of the results of the method of the present invention and existing methods. Detailed Implementation

[0058] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0059] Example 1

[0060] This invention relates to a multi-stage interactive low-light image enhancement method based on dual-channel normalized streams, the specific steps of which are as follows:

[0061] Step 1: Convert the low-light RGB image to the YUV color space to obtain the luminance channel image Y and the color channel image U and color channel image V;

[0062] Step 2: Construct the overall network and train the overall network. Input the luminance channel image Y and color channel images U and V obtained in Step 1 into the trained overall network to obtain the corrected luminance channel image Y″, color channel image U″, and color channel image V″.

[0063] Step 3: Convert the corrected luminance channel image Y″, color channel image U″, and color channel V″ obtained in Step 2 into RGB images to obtain an enhanced RGB image with natural brightness and color performance, while maintaining the same texture structure as the original. Figure 1 To.

[0064] Example 2

[0065] Based on Example 1, the overall network constructed in step 2 consists of a brightness perception enhancement network, a dual-channel Normalizing Flow module, and a joint correction and reconstruction network;

[0066] The luminance channel image Y is used as the input to the luminance perception enhancement network, and the output is an enhanced luminance channel image Y′.

[0067] Color channel image U and color channel image V are used as a dual-channel Normalizing Flow module to output an enhanced color channel image U′ and an enhanced color channel image V′.

[0068] The enhanced luminance channel image Y′, enhanced color channel image U′, and enhanced color channel V′ are used as inputs to the joint correction and reconstruction network, which outputs the corrected luminance channel image Y″, color channel image U″, and color channel V″.

[0069] Example 3

[0070] Based on Example 2, such as Figure 2As shown, the brightness perception enhancement network is composed of a multi-scale feature extraction module, a multi-scale feature fusion module, a brightness perception attention module, and a transposed convolution recovery module. The brightness perception enhancement network can capture brightness change details at different resolution levels, highlight significant areas of the image, enhance local contrast, and effectively improve the visual perception quality of low-light images.

[0071] The processing procedure of the brightness perception enhancement network is as follows: the brightness channel image Y is input into the multi-scale feature extraction module and the brightness perception attention module, the output of the multi-scale feature extraction module is input into the multi-scale feature fusion module, the multi-scale feature fusion module outputs a multi-scale feature map, the brightness perception attention module outputs a brightness perception weight map, the multi-scale feature map and the brightness perception weight map are multiplied element by element to obtain a weighted brightness feature map, and the weighted brightness feature map is input into the transposed convolution recovery module to obtain the enhanced brightness channel image Y'.

[0072] The multi-scale feature extraction module is composed of three convolution extraction modules, each of which is composed of a convolution layer, a Batch Normalization layer, and a ReLU activation function in sequence. The convolution kernel size of the convolution layer in the first convolution extraction module is 3x3, the channel number is 32, and the feature map F1 output by the first convolution extraction module has a size of HxWx32. The convolution kernel size of the convolution layer in the second convolution extraction module is 3x3, the channel number is 64, and the feature map F2 output by the second convolution extraction module has a size of H / 2xW / 2x64. The convolution kernel size of the convolution layer in the third convolution extraction module is 3x3, the channel number is 128, and the feature map F3 output by the third convolution extraction module has a size of H / 4xW / 4x128. The multi-scale feature extraction module is used to extract image features of different semantic levels from the input brightness channel Y from shallow to deep.

[0073] The multi-scale feature fusion module includes a spatial alignment stage and a channel alignment stage. In the spatial alignment stage, the feature map F1 remains unchanged, the feature map F2 is upsampled by a transposed convolution with a convolution kernel size of 4x4, a step size of 2, and a padding of 1, and the feature map size becomes HxWx64, and the feature map F3 is upsampled by a transposed convolution with a convolution kernel size of 8x8, a step size of 4, and a padding of 2, and the feature map size becomes HxWx128. In the channel alignment stage, the HxWx32 feature map, the HxWx64 feature map, and the HxWx128 feature map processed in the spatial alignment stage are respectively compressed in the channel dimension by a 1x1 convolution to unify the channel number to 32 to obtain the feature map F1', the feature map F2', and the feature map F3'. The feature map F1', the feature map F2', and the feature map F3' are spliced in the channel dimension, and then a 1x1 convolution is used to compress the channel to 32.

[0074] The luminance-aware attention module is sequentially composed of a global average pooling and an SE (Squeeze-and-Excitation) module; the luminance-aware attention module is used to give different weights to different regions, highlight the details of the weak light region, and suppress irrelevant regions.

[0075] The transposed convolution restoration module is composed of two layers of transposed convolution modules, and each layer of transposed convolution module is sequentially composed of a transposed convolution layer, a ReLU activation, and a normalization layer; the convolution kernel of the transposed convolution in each layer of transposed convolution module is 4x4, the step is 2, and the padding is 1; the input channel number of the first layer of convolution module is 128, the output channel number is 64, the input channel number of the second layer of convolution module is 64, and the output channel number is 32.

[0076] Embodiment 4

[0077] On the basis of embodiment 3, as shown in Figure 3 The double-channel Normalizing Flow module is composed of a U-channel Flow module, a V-channel Flow module, and two multi-stage feature information interaction modules; the double-channel Normalizing Flow module accurately models the non-Gaussian distribution of the color channel in the weak light image, can realize accurate mapping and inverse mapping between the color channel image U and V and the standard Gaussian distribution, and thus enhances the naturalness and consistency of the image color;

[0078] The U-channel Flow module is sequentially composed of Stage1, Stage2, and Stage3.

[0079] The V-channel Flow module is sequentially composed of Stage4, Stage5, and Stage6.

[0080] Stage1, Stage2, Stage3, Stage4, Stage5, and Stage6 are all composed of a learnable normalization layer, a reversible 1x1 convolution layer, and an affine coupling layer; the reversible 1x1 convolution layer is composed of a convolution layer with a convolution kernel of 1x1 (where the weight matrix is a square matrix and is non-singular); the affine coupling layer divides the input into A and B, A remains unchanged, B is scaled and translated through a lightweight network to output B', A and B' are spliced to obtain the final output; the learnable normalization layer normalizes the mean and variance of each channel of the input to ensure stable data distribution and facilitate subsequent learning; the reversible 1x1 convolution layer is used for linear mixing of the input, thereby increasing the dependency between each color channel image; this operation is completely reversible, and only the weight needs to be inverted when inverting;

[0081] Each multi-stage feature information interaction module is sequentially composed of a feature concatenation layer, a multi-head attention module and a residual connection.

[0082] The processing procedure of the first multi-stage feature information interaction module is as follows: the outputs of Stage 1 in the U-channel Flow module and Stage 4 in the V-channel Flow module are concatenated by channels to form a joint semantic tensor, the joint semantic tensor is input into the feature concatenation layer for dimension reduction to obtain a reduced feature, the reduced feature is input into the multi-head attention module, the output of the multi-head attention module is connected with the joint semantic tensor in a residual manner to obtain a final output; the processing procedure of the second multi-stage feature information interaction module is the same as that of the first multi-stage feature information interaction module, and the input of the second multi-stage feature information interaction module is a joint semantic tensor formed by concatenating the outputs of Stage 2 in the U-channel Flow module and Stage 5 in the V-channel Flow module by channels;

[0083] Each multi-stage feature information interaction module is used to improve the consistency of color correction and suppress the problems of channel deviation and mismatch.

[0084] The feature concatenation layer is composed of at least one convolution layer, the convolution kernel of each convolution layer is 1x1, the step is 1 and the padding is 0; the multi-head attention module is 4-head attention, the query is provided by the color channel image U, the key and the value are provided by the color channel image V; the residual connection is sequentially composed of a convolution with a convolution kernel of 1x1 and a skip connection; each convolution layer is used to reduce the concatenated intermediate features output at the end of U and V stages to a specified embedding space to form a joint semantic representation; the multi-head attention module is used to learn the dependency relationship between the semantic regions of U and V, and realize the complementation and alignment of cross-channel semantic information between U and V channels; the residual connection is used to maintain the input semantic information and form a bidirectional information flow at the same time, and this structure can explicitly make U and V jointly learn the semantic information of the other party;

[0085] The processing procedure of the dual-channel Normalizing Flow module is as follows: the color channel image U is input into Stage 1, the color channel image V is input into Stage 4, the output features of Stage 1 and the output features of Stage 4 are input into the first multi-stage feature information interaction module, the output of the first multi-stage feature information interaction module is input into Stage 2 and Stage 5 respectively, the output features of Stage 2 and the output features of Stage 5 are input into the second multi-stage feature information interaction module, the output of the second multi-stage feature information interaction module is input into Stage 3 and Stage 6 respectively, Stage 3 outputs the enhanced color channel image U', and Stage 6 outputs the enhanced color channel image V'.

[0086] Embodiment 5

[0087] Based on the embodiment 4, as shown in Figure 4 the joint correction and reconstruction network is composed of an encoder and a decoder, and the encoder and the decoder are connected in a skip manner; the enhanced luminance channel image Y', the enhanced color channel image U', and the enhanced color channel image V' are input into the joint correction and reconstruction network, and the corrected luminance channel image Y", the corrected color channel image U", and the corrected color channel image V" are output; the joint correction and reconstruction network integrates and optimizes the three-channel information, and further improves the overall structural consistency and color continuity of the image; the joint correction and reconstruction network is constructed based on U-Net, and low-level texture information is retained through the skip connection;

[0088] The encoder is composed of three convolution modules, and each convolution module is composed of a convolution layer, a normalization layer, and a ReLU activation function in sequence; the convolution kernel of the convolution layer in the first convolution module is 3x3, the step is 2, the input channel number is 32, and the output channel number is 64; the convolution kernel of the convolution layer in the second convolution module is 3x3, the step is 2, the input channel number is 64, and the output channel number is 128; the convolution kernel of the convolution layer in the third convolution module is 3x3, the step is 2, the input channel number is 128, and the output channel number is 256.

[0089] The decoder is composed of three transposed convolution modules and a Sigmoid activation function in sequence; each transposed convolution module is composed of a transposed convolution layer and a ReLU activation function in sequence; the convolution kernel of the transposed convolution in the first transposed convolution module is 4x4, the step is 2, the input channel is 256, and the output channel number is 128; the convolution kernel of the transposed convolution in the second transposed convolution module is 4x4, the step is 2, the input channel is 128, and the output channel number is 64; the convolution kernel of the transposed convolution in the third transposed convolution module is 4x4, the step is 2, the input channel is 64, and the output channel number is 32.

[0090] Each layer of the convolution module in the encoder is connected with the corresponding transposed convolution module in the decoder, i.e., the first layer of the convolution module in the encoder is connected with the third transposed convolution module in the decoder, the second layer of the convolution module in the encoder is connected with the second transposed convolution module in the decoder, and the third layer of the convolution module in the encoder is connected with the first transposed convolution module in the decoder.

[0091] Embodiment 6

[0092] On the basis of embodiment 5, the total loss function of the overall network is:

[0093] (2)

[0094] In formula (2), is a degree reconstruction loss, which is used to supervise the training of the brightness perception enhancement network, so that the enhanced brightness channel image is close to the real brightness image in structure and value ; is a maximum likelihood loss, which is used to supervise the U and V channel Normalizing Flow module, and ensure that the two-channel Normalizing Flow module learns an accurate reversible mapping of the channel distribution; is a color consistency loss, which is used to control the enhanced U / V channel to be close to the real natural color in value and distribution, and prevent problems such as color deviation and saturation distortion; is a joint perception loss;

[0095] wherein,

[0096] (3)

[0097] In formula (3), is a structural similarity index, which is used to measure the structural fidelity of an image; and are loss weight coefficients; is a real brightness image;

[0098] (4)

[0099] In formula (4), is the probability density of the Flow module after shooting under the standard Gaussian distribution, is a negative log-likelihood;

[0100] (5)

[0101] In formula (5), and are real color channel images; To enhance the color distribution of the color channel image U′, For images under normal lighting conditions The color distribution of the channels is used to make the color distribution of the enhanced color channel image U′ more natural. KL divergence measures the distance between two probability distributions; , and These are the loss weighting coefficients;

[0102] (6)

[0103] In equation (6), For YUV space transformation functions, For the enhanced RGB image, This is a normal light image. Indicates the perceived loss weight;

[0104] The overall network training adopts an end-to-end joint training approach. The brightness perception enhancement network, the dual-channel Normalizing Flow module, and the joint correction and reconstruction network are simultaneously optimized along the same gradient backpropagation path. The training set uses the publicly available LOL-v2 low-light image dataset, including approximately 1000 pairs of low-light images and reference images. The images are resized to 512×512 pixels, normalized (0~1), and uniformly converted to YUV format before being input into the network. The hyperparameters are set as follows: =1.0, =0.1, =0.3, =0.05, learning rate set to 0.0001, batch size=8, training iterations 300 epochs. During training, the network converged stably without gradient explosion or oscillation issues.

[0105] like Figure 5 As shown in the figure, the method of the present invention is compared with the existing GT method. It can be seen from the figure that the enhancement effect of the method of the present invention is better than that of the GT method.

Claims

1. A multi-stage information-interactive low-light image enhancement method based on dual-channel normalized stream, characterized in that, The specific steps are as follows: Step 1: Convert the low-light RGB image to the YUV color space to obtain the luminance channel image Y, the color channel image U, and the color channel image V; Step 2: Construct the overall network by inputting the luminance channel image Y and color channel images U and V obtained in Step 1 into the overall network to obtain the corrected luminance channel image Y″, color channel image U″, and color channel image V″. In step 2, the overall network consists of a brightness perception enhancement network, a dual-channel Normalizing Flow module, and a joint correction and reconstruction network; The luminance channel image Y is used as the input to the luminance perception enhancement network, and the output is an enhanced luminance channel image Y′. Color channel image U and color channel image V are used as inputs to the dual-channel Normalizing Flow module, and the outputs are enhanced color channel image U′ and enhanced color channel image V′. The enhanced luminance channel image Y′, enhanced color channel image U′, and enhanced color channel image V′ are used as inputs to the joint correction and reconstruction network, and the outputs are the corrected luminance channel image Y″, corrected color channel image U″, and corrected color channel image V″. The joint correction and reconstruction network consists of an encoder, a decoder, and skip connections; Step 3: Convert the corrected luminance channel image Y″, color channel image U″, and color channel image V″ obtained in Step 2 into RGB images to obtain the enhanced RGB image.

2. The multi-stage interactive low-light image enhancement method based on dual-channel normalized stream according to claim 1, characterized in that, The brightness perception enhancement network consists of a multi-scale feature extraction module, a multi-scale feature fusion module, a brightness perception attention module, and a transposed convolution recovery module; The processing procedure of the brightness perception enhancement network is as follows: The brightness channel image Y is input into the multi-scale feature extraction module and the brightness perception attention module respectively. The output of the multi-scale feature extraction module is input into the multi-scale feature fusion module. The multi-scale feature fusion module outputs a multi-scale feature map, and the brightness perception attention module outputs a brightness perception weight map. The multi-scale feature map and the brightness perception weight map are multiplied element-wise to obtain a weighted brightness feature map. The weighted brightness feature map is input into the transposed convolution recovery module to output the enhanced brightness channel image Y′.

3. The multi-stage interactive low-light image enhancement method based on dual-channel normalized flow according to claim 2, characterized in that, The multi-scale feature extraction module consists of at least three convolutional extraction modules, each of which consists of a convolutional layer, a normalization layer, and a ReLU activation function. The multi-scale feature fusion module includes a spatial alignment stage and a channel alignment stage; The brightness-sensing attention module consists of global average pooling and an SE module, respectively. The transposed convolution recovery module consists of at least two transposed convolution modules, each of which consists of a transposed convolution layer, a ReLU activation layer, and a normalization layer in sequence.

4. The multi-stage interactive low-light image enhancement method based on dual-channel normalized stream according to claim 1, characterized in that, The dual-channel Normalizing Flow module consists of a U-channel Flow module, a V-channel Flow module, and two multi-stage feature information interaction modules; The U-channel Flow module consists of Stage1, Stage2, and Stage3 in sequence; The V-channel Flow module consists of Stage 4, Stage 5, and Stage 6 in sequence; Stage 1, Stage 2, Stage 3, Stage 4, Stage 5, and Stage 6 are all composed of learnable normalization layers, reversible 1×1 convolutional layers, and affine coupling layers; The affine coupling layer splits the input in half into A and B. A remains unchanged, while B passes through a lightweight network to output B′. A and B′ are then concatenated to obtain the final output. Each multi-stage feature information interaction module consists of a feature splicing layer, a multi-head self-attention module, and a residual connection in sequence. The feature splicing layer consists of at least one convolutional layer; The residual connections consist of convolutions with 1×1 kernels and skip connections.

5. The multi-stage interactive low-light image enhancement method based on dual-channel normalized flow according to claim 4, characterized in that, The processing procedure of the dual-channel Normalizing Flow module is as follows: the color channel image U is input into Stage 1, and the color channel image V is input into Stage 4. The output features of Stage 1 and Stage 4 are input together into the first multi-stage feature information interaction module. The output of the first multi-stage feature information interaction module is input into Stage 2 and Stage 5 respectively. The output features of Stage 2 and Stage 5 are input together into the second multi-stage feature information interaction module. The output of the second multi-stage feature information interaction module is input into Stage 3 and Stage 6 respectively. Stage 3 outputs the enhanced color channel image U′, and Stage 6 outputs the enhanced color channel image V′.

6. The multi-stage information-interactive low-light image enhancement method based on dual-channel normalized stream according to claim 1, characterized in that, The encoder consists of three convolutional modules, each of which consists of a convolutional layer, a normalization layer, and a ReLU activation function. The decoder consists of three transposed convolutional modules and a Sigmoid activation function; each transposed convolutional module consists of a transposed convolutional layer and a ReLU activation function. Each convolutional module in the encoder is skipped to the corresponding transposed convolutional module in the decoder.

7. The multi-stage interactive low-light image enhancement method based on dual-channel normalized stream according to claim 1, characterized in that, In step 2, the overall loss function of the network is: （2） In equation (2), For brightness reconstruction loss; The maximum likelihood loss for Flow; For loss of color consistency; For joint perception of loss; in, （3） In equation (3), It is a structural similarity index used to measure the structural fidelity of an image; and These are the loss weighting coefficients; A true brightness image; （4） In equation (4), This represents the probability density of the Flow module under a standard Gaussian distribution after mapping. The likelihood is negative log-likelihood. （5） In equation (5), and This is a true color channel map; To enhance the color distribution of the color channel image U′, For images under normal lighting Color distribution of the channel, KL divergence measures the distance between two probability distributions; , and These are the loss weighting coefficients; （6） In equation (6), For YUV space transformation functions, For the enhanced RGB image, This is a normal light image. This represents the perceived loss weight.

Images