A low-light image enhancement method using time-frequency domain hierarchical processing

Through the combination of time-frequency domain hierarchical processing and deep learning framework, the problems of brightness improvement and noise suppression in low-light image enhancement are solved, and natural and clear images are generated, suitable for multiple fields, improving image quality and the accuracy of visual tasks .

CN114219722BActive Publication Date: 2024-11-19SHENZHEN INTELLECTUAL PROPERTY OPERATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111336761.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-12
Publication Date
2024-11-19
Estimated Expiration
2041-11-12

AI Technical Summary

Technical Problem

The existing low-light image enhancement algorithm cannot effectively improve image brightness and suppress noise, resulting in unnatural and noisy images, affecting the visual perception and performance of computer vision systems.

Method used

Using the low-light image enhancement method of time-frequency domain hierarchical processing, combined with the deep learning framework, a network framework is built, and through time-domain brightness enhancement, frequency-domain noise suppression and time-domain detail reconstruction modules, image brightness enhancement and noise suppression are achieved, and natural , clear result images.

Benefits of technology

It achieves natural brightness improvement and noise suppression for low-light images, and is efficient, robust and versatile. It is suitable for multiple fields such as object detection and autonomous driving, improving image quality and accuracy of visual tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114219722B_ABST
    Figure CN114219722B_ABST
Patent Text Reader

Abstract

The present invention relates to a low-light image enhancement method using time-frequency domain hierarchical processing, including the following steps: S1, constructing a low-light image data set, including the collection and annotation of the low-light image data set, and reasonably dividing the training data set and the test data set; S2, establishing a low-light image enhancement network framework, and using the convolutional neural network in deep learning to build the entire network framework; S3, performing model training, and using the training data set divided in step S1 to iteratively train the low-light image enhancement network framework constructed in step S2; S4, performing model testing, and using the model saved after training in step S3 to detect the test data set divided in step S1. Under the deep learning framework, the present invention combines the time domain and the frequency domain to construct a network framework, and uses a data-driven method to hierarchically process low-light images, realize image brightness stretching, and generate clear, low-noise result images, making the low-light image enhancement task more efficient.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of low-light image enhancement, and in particular to a low-light image enhancement method using time-frequency domain hierarchical processing. Background Art

[0002] With the gradual improvement of living standards and the development of high-tech technology, people often record the details of life by taking photos. At the same time, in industry, intelligent devices efficiently process visual tasks by capturing various images, achieving higher efficiency than manual operations. Therefore, images have become an indispensable part of human life. However, due to low-light environments and limited and expensive camera equipment, people often capture some low-quality images, such as low-light images. Such low-light images usually have disadvantages such as low brightness, low contrast, and high noise, which not only affects the beauty of the image and human visual perception, but also challenges many intelligent systems that rely on computer vision algorithms, such as all-day autonomous driving and biometric recognition, target detection, etc., resulting in limited performance.

[0003] In order to effectively improve the quality of low-light images, many low-light image enhancement algorithms have emerged. These algorithms can be divided into three categories: methods based on traditional theories, methods based on the retinal cortex (Retinex), and methods based on deep learning. The traditional classic method is histogram equalization. This method counts the grayscale values ​​of pixels according to the image histogram, and then performs pixel equalization on them, so that the brightness is improved. However, this method does not consider the relationship between neighboring pixels and only processes the image pixels in a single way, so the resulting image often looks unnatural. Retinex theory believes that any image can be decomposed into a reflection component that reflects the intrinsic properties of the object and an illumination component affected by the environment. The brightness of external objects perceived by the human eye is usually determined by the reflection component of the object, but because the decomposition process of the reflection component and the illumination component is an ill-posed problem, it is difficult for the Retinex-based method to completely separate the two components and effectively enhance them. The deep learning-based method captures natural image statistical information from large-scale data, flexibly learns, has certain applicability to real images, has high generalization ability, and is low cost. Therefore, a technical solution for image brightness enhancement and noise suppression can be further developed through the framework of deep learning. Summary of the invention

[0004] In order to solve the technical problems existing in the prior art, the present invention provides a low-light image enhancement method using time-frequency domain hierarchical processing. Under the deep learning framework, a network framework is constructed by combining the time domain and the frequency domain, and the low-light image is hierarchically processed in a data-driven manner. While stretching the image brightness range, a natural, clear, and low-noise result image is generated, making the low-light image enhancement task more efficient.

[0005] The present invention is implemented by the following technical solution: A low-light image enhancement method using time-frequency domain hierarchical processing, comprising the following steps:

[0006] S1. Construct a low-light image dataset, including the collection and annotation of the low-light image dataset, and reasonably divide the training dataset and the test dataset;

[0007] S2. Establish a low-light image enhancement network framework. Use the convolutional neural network in deep learning to build the entire network framework. Design corresponding modules for the problems existing in low-light images and build a low-light image enhancement network based on deep learning.

[0008] S3, performing model training, using the training data set divided in step S1 to iteratively train the low-light image enhancement network framework constructed in step S2;

[0009] S4: Perform model testing, using the model saved after training in step S3 to test the test data set divided in step S1.

[0010] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0011] 1. The present invention obtains a universal low-light image enhancement network model by training a dataset composed of a mixture of natural low-light images and artificially synthesized low-light images. The model can not only process low-light images produced by natural scenes, but also take into account more complex lighting conditions. It can enhance the brightness well while significantly suppressing noise and restoring details. It has high robustness and versatility.

[0012] 2. The network model constructed by the present invention processes images based on deep learning technology, with fast speed and natural and clear output results.

[0013] 3. The network constructed by the present invention is trained end-to-end, has low requirements for operating equipment, low cost, and small resource occupation. Therefore, it has obvious application value in many fields such as target detection, digital monitoring, and autonomous driving. It can not only improve the quality of life, but also further improve the accuracy of visual tasks and promote the development of artificial intelligence technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the method of the present invention;

[0015] Figure 2 This is a low-light image enhancement network framework diagram of the present invention;

[0016] Figure 3 It is a network structure diagram of the dynamic enhancement unit of the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0018] Example

[0019] like Figure 1 As shown, the present embodiment provides a low-light image enhancement method using time-frequency domain hierarchical processing, including the following steps:

[0020] S1. Construct a low-light image dataset, including the collection and annotation of the low-light image dataset, and reasonably divide the training dataset and the test dataset;

[0021] S2. Establish a low-light image enhancement network framework. Use the convolutional neural network in deep learning to build the entire network framework. Design corresponding modules for the problems existing in low-light images and build a low-light image enhancement network based on deep learning.

[0022] S3, performing model training, using the training data set divided in step S1 to iteratively train the low-light image enhancement network framework constructed in step S2;

[0023] S4: Perform model testing, using the model saved after training in step S3 to test the test data set divided in step S1.

[0024] In this embodiment, the specific steps of constructing the low-light image dataset in step S1 are as follows:

[0025] S11. Collect images and obtain low-light images through two ways, including collecting low-light images from various photography websites and using existing SLR cameras and mobile phones of various brands to manually collect images in real daytime and nighttime scenes. Specifically, in daytime scenes, by adjusting the aperture of the camera, two types of images of the same scene can be collected, namely, normal light images and low light images. At night and in low-light environments, the aperture of the camera is set to a normal range, and the object is directly photographed to collect low-light images.

[0026] S12, labeling the data set, classifying the obtained images, manually labeling the images collected from the website and taken at night, calling the low-light image and the corresponding normal-light image of the same scene a set of image pairs, and renaming the file names of the image pairs in natural number order;

[0027] S13, divide and label the data set, and randomly divide all the collected images into a training data set and a test data set in a ratio of 7:3.

[0028] like Figure 2As shown, in this embodiment, the low-light image enhancement network framework in step S2 includes:

[0029] Temporal brightness enhancement module, used to improve the brightness level of dark light images;

[0030] Frequency domain noise suppression module, used to filter out image noise;

[0031] Time domain detail reconstruction module: Due to the filtering effect of the frequency domain noise suppression module, the image becomes smooth and the details are lost. Therefore, the time domain detail reconstruction module is mainly used to restore the structure, texture and other details of the image;

[0032] Each module is composed of a dynamic enhancement unit (DEU) as the basic element, using different construction methods, and uses L1 regularization loss, total variation difference (TV) loss and Laplace loss as a set of constraints used as the total loss function of the network.

[0033] Specifically, in this embodiment, the specific implementation process of the time domain brightness enhancement module is as follows:

[0034] S211, the input image is subjected to a convolution operation with a kernel of 3×3 to obtain a set of features;

[0035] S212, inputting the feature into a DEU enhanced feature that works in a human eye-like mechanism, so that it can better characterize the brightness level of the image and enhance the brightness using statistical characteristics;

[0036] S213, using a layer of convolution with a kernel of 3×3 and Tanh activation operation to encode the image and output it to complete the preliminary brightness enhancement.

[0037] Specifically, in this embodiment, the specific implementation process of the frequency domain noise suppression module is as follows:

[0038] S221, transforming the image features obtained by the time domain brightness enhancement module into the frequency domain using a wavelet transform (Discrete Wavelet Transform, DWT);

[0039] S222, using DEU and basic convolution to process features;

[0040] S223, the processed image is gradually transformed into a time domain image by performing Discrete Inverse Wavelet Transform (IDWT), and then enters the processing of the time domain detail reconstruction module.

[0041] Specifically, in this embodiment, the specific implementation process of the time domain detail reconstruction module is as follows:

[0042] S231, cascading the output result graphs of the time domain brightness enhancement module and the frequency domain noise suppression module, and inputting them into the time domain detail reconstruction module;

[0043] S232, the image is passed through several convolutional layers and multiple dynamic enhancement units DEU to reconstruct detail information such as image edge enhancement and color restoration;

[0044] S233, output the final enhanced image of the entire network.

[0045] In this embodiment, the specific process of step S211 is: the input image S∈R with an image size of H×W and a channel number of 3 H×W×3 , a convolutional layer with a kernel of 3×3 and a leaky rectified linear unit (LReLU) activation operation are used to extract a set of features of size [8, 64, 256, 256], which contains information such as image brightness, color, and details.

[0046] like Figure 3 As shown, in this embodiment, the dynamic enhancement unit DEU in step S212 mainly adopts a retina-like working processing mechanism to further process the input features in a parallel channel manner. For the input visual signal, the human retina uses two types of cells to work together, namely, dwarf cells and parasol cells. Among them, the dwarf cells have a small receptive field and are used to encode high-frequency information, such as details and edges in the scene; while parasol cells are good at processing low-frequency parts in visual information, including the global structure of the scene, and its main feature is a large receptive field. Therefore, according to this principle, the dynamic enhancement unit DEU uses convolutions with kernels of 3×3 and 5×5 to represent dwarf cells and parasol cells to form two processing pathways to enhance the features. These two parallel channels can extract feature information of different receptive fields. Then, the output features of the two pathways are fused using feature addition operations. This process simulates the working mechanism of the optic chiasm nerve of the human brain. Different information can be aggregated through addition operations, so that the features can better represent the image. Then, the fused features are respectively input into the network pathway composed of convolution layers in a parallel manner to enhance the learned features and fuse their information in a cascade manner. Furthermore, a set of convolutions with kernels of 1×1 and 3×3 are used again to balance the fused features to make them smoother and the information exchange more sufficient. The whole process can be quantitatively expressed as:

[0047] f = Rconv3(X deu )+Rconv5(X deu ) (1)

[0048]

[0049]

[0050] Among them, X deu represents the input features of the dynamic enhancement unit DEU; RconvN(.) is a sequential operation combining a convolution operation with a kernel of N×N and an LReLU activation operation; F dam (·) is the Dynamic Attention Module (DAM) operation; It is a cascade operation; is a pixel-by-pixel addition operation; f is the intermediate output; f deu It is the output result of the final dynamic enhancement unit DEU.

[0051] Specifically, F dam (·) simulates the selective working mechanism of the human cerebral cortex. After humans capture external scene information, it is transmitted to the cerebral cortex through the eyes, and the control center brain center will filter out useful information according to needs and discard irrelevant parts. Therefore, according to this mechanism, the dynamic enhancement unit DEU is followed by a dynamic attention mechanism module DAM, which is used to filter features so that the network pays attention to the parts that are beneficial to the task. For example, for the brightness enhancement task of the time domain brightness enhancement module, it pays more attention to the brightness information; the frequency domain noise suppression module mainly acts on the noise area; and the time domain detail reconstruction module needs to pay attention to the edge structure details of the image. The structure of the dynamic attention mechanism module DAM is as follows Figure 3 As shown in the figure, its structure consists of three parallel branches, which are channel attention branch, filtering branch, and pixel attention branch in order from top to bottom. Specifically, the channel attention branch consists of global average pooling, 1×1 convolution, ReLU activation function, 1×1 convolution, and Sigmoid activation operations in sequence; compared with the channel attention branch, the pixel attention branch removes the global average pooling operation, so that its processing level is the image pixel point. For the filtering branch, the classification processing system is mainly composed of linear layers and Softmax operations. The output result of this branch is a set of classification score values ​​(x, y). Therefore, the working process of the dynamic attention mechanism module DAM can be expressed as:

[0052]

[0053] Among them, X dam It is the input feature of the dynamic attention mechanism module DAM; Represents the channel attention branch of the dynamic attention mechanism module DAM; represents the pixel attention branch of the dynamic attention mechanism module DAM; x, y are the output scores of the screening branch; f dam is the output feature.

[0054] In this embodiment, the preliminary brightness enhancement process in step S213 can be expressed as:

[0055] S l =Tanh(Rconv3(F deu (Rconv3(S)))) (5)

[0056] Where S represents the input image; F deu (·) is the operation of the dynamic enhancement unit DEU; Tanh(·) is the hyperbolic tangent activation function; S l Represents the output feature, which is the output image after brightness enhancement.

[0057] In this embodiment, the wavelet transform DWT in step S221 mainly filters each row and column of the image through a one-dimensional filter function and a wavelet function, and downsamples in the horizontal and vertical directions. Thus, four sub-images I can be calculated. LL , I LH , I HL , I HH For the input image S(x, y), the four sub-images output by the two-dimensional wavelet transform can be expressed as follows:

[0058] I LL =φ(x)φ(y) (6)

[0059]

[0060]

[0061]

[0062] Among them, x is the horizontal pixel value of the image; y is the vertical pixel value; φ(·) and are the one-dimensional scaling function and the wavelet function respectively; I LL It is the low-frequency component of the image, containing basic scene information, such as color, brightness, content, etc.; I LH , I HL , I HH The high-frequency components retain the image details, such as noise and edges, after low-pass filtering. In contrast, the inverse wavelet transform uses the same filter and upsampling method to merge the four sub-images into the original image. Since the wavelet transform DWT and the inverse wavelet transform IDWT involve downsampling and upsampling, the image size will be reduced to one-fourth of the original size after each DWT operation, and the number of channels will increase fourfold accordingly; on the contrary, the inverse wavelet transform IDWT will quadruple the image size and reduce the number of channels by one-fourth.

[0063] like Figure 2As shown, in this embodiment, in the frequency domain noise suppression module, after the image is sequentially subjected to wavelet transform DWT, convolution layer and dynamic enhancement unit DEU operations, features {x1, x2, x3} can be obtained in turn, and their sizes are {256, 128, 64} respectively. x3 is first input into the convolution layer and dynamic enhancement unit DEU in a bottom-up order, and the noise therein can be suppressed for the first time. Then, it is upsampled to the same size as x2 through inverse wavelet transform IDWT and cascaded with it, and then the convolution layer and dynamic enhancement unit DEU are used to perform a second noise suppression on it. Similarly, x2 is inversely transformed with wavelet and cascaded with x1 as the input of the third noise suppression operation. After processing, the image size is finally obtained, which is the same as the original image. Figure 1 The enhanced image S d .

[0064] In this embodiment, after the feature enters the time domain detail reconstruction module, it undergoes 1×1 convolution, 3×3 convolution, three dynamic enhancement units DEU and 3×3 convolution operations in sequence. Although the network has only 5 layers, the details can be well restored under the drive of the loss function. Therefore, the output result S of the time domain detail reconstruction module is s This is the final enhanced image of the entire network.

[0065] Specifically, the loss function is defined as follows:

[0066] L total =l1+l2+l3 (10)

[0067] l1=||S l -S normal ||1 (11)

[0068] l2=||S d -S normal ||1+||ΔS d ||1 (12)

[0069] l3=||S s -S normal ||1+||F lap (s s )-F lap (S normal )||1 (13)

[0070] Among them, L total represents the total loss of the low-light enhancement network; l1, l2, and l3 are the losses of the temporal brightness enhancement module, the frequency domain noise suppression module, and the temporal detail reconstruction module, respectively; S l , S d , S sThey are the output result diagrams of the time domain brightness enhancement module, the frequency domain noise suppression module, and the time domain detail reconstruction module; S normoal is the label image, i.e. the corresponding normal illumination image; ‖·‖1 is the L1 regularization function; Δ represents the sum of the horizontal and vertical gradients of the image; F lap (· is the Laplace transform function.

[0071] In this embodiment, the specific process of step S3 is as follows:

[0072] S31. When training the model, the input image size is adjusted to 256×256. The network is trained using the adaptive moment estimation optimizer with a running average coefficient of (0.9, 0.999) and a weight decay of 0.000008. The number of training iterations is 150, the batch size is 8, the initial learning rate is 0.0002, and the learning plan strategy is adopted. After the number of iterations reaches 60, the learning rate decreases by 10%;

[0073] S32. Save the model after every 10 training iterations.

[0074] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A low-light image enhancement method using time-frequency domain hierarchical processing, characterized in that: The following steps are involved: S1. Construct a low-light image dataset, including the collection and annotation of the low-light image dataset, and reasonably divide the training dataset and the test dataset; S2. Establish a low-light image enhancement network framework. Use the convolutional neural network in deep learning to build the entire network framework. Design corresponding modules for the problems existing in low-light images and build a low-light image enhancement network based on deep learning. S3, performing model training, using the training data set divided in step S1 to iteratively train the low-light image enhancement network framework constructed in step S2; S4, perform model testing, using the model saved after training in step S3 to test the test data set divided in step S1; The low-light image enhancement network framework in step S2 includes: Temporal brightness enhancement module, used to improve the brightness level of dark light images; Frequency domain noise suppression module, used to filter out image noise; The time domain detail reconstruction module is used to restore the details of the image, including the structure and texture of the image; Each module is composed of a dynamic enhancement unit DEU as a basic element, using a variety of construction methods, and uses L1 regularization loss, total variation difference TV loss and Laplace loss as a set of constraints used as the total loss function of the network; In step S212, the dynamic enhancement unit DEU uses two processing paths with convolution kernels of 3×3 and 5×5 to enhance the features, and uses these two processing paths to extract feature information; then the output features of the two paths are fused by feature addition operation, and the fused features are input into the network path composed of convolution layers in parallel, the learned features are enhanced and their information is fused in a cascade manner, and then a group of convolution kernels of 1×1 and 3×3 are added to balance the fused features. The whole process can be quantitatively expressed as: f=Rconv3(X deu )+Rconv5(X deu ) (1) Among them, X deu represents the input features of the dynamic enhancement unit DEU; RconvN(.) is a sequential operation combining a convolution operation with a kernel of N×N and an LReLU activation operation; F dam (·) is the dynamic attention mechanism module DAM operation; It is a cascade operation; is a pixel-by-pixel addition operation; f is the intermediate output; f deu It is the final output result of the dynamic enhancement unit DEU; The dynamic attention mechanism module DAM consists of three parallel branches, which are channel attention branch, filter branch, and pixel attention branch in order from top to bottom; the channel attention branch consists of global average pooling, 1×1 convolution, ReLU activation function, 1×1 convolution, and Sigmoid activation operations in sequence; the filter branch, by using the linear layer and Softmax operation to form a classification processing system, the output of this branch is a set of classification score values ​​(x, y); the pixel attention branch consists of 1×1 convolution, ReLU activation function, 1×1 convolution, and Sigmoid activation operations in sequence; the working process of the dynamic attention mechanism module DAM can be expressed as: Among them, X dam It is the input feature of the dynamic attention mechanism module DAM; Represents the channel attention branch of the dynamic attention mechanism module DAM; represents the pixel attention branch of the dynamic attention mechanism module DAM; x, y are the output scores of the screening branch; f dam is the output feature.

2. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 1, characterized in that: The specific steps of constructing the low-light image dataset in step S1 are as follows: S11. Collect images and obtain low-light images through two ways, including collecting low-light images from various photography websites and using existing SLR cameras and mobile phones of various brands to manually collect images in real daytime and nighttime scenes; in daytime scenes, adjust the aperture of the camera to collect two types of images of the same scene, namely, normal light images and low light images; in night and low-light environments, set the aperture of the camera to the normal range, shoot objects, and collect low-light images; S12, labeling the data set, classifying the obtained images, manually labeling the images collected from the website and taken at night, calling the low-light image and the corresponding normal-light image of the same scene a set of image pairs, and renaming the file names of the image pairs in natural number order; S13, divide and label the data set, and randomly divide all the collected images into a training data set and a test data set in a ratio of 7:

3.

3. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 1, characterized in that: The specific implementation process of the time domain brightness enhancement module is as follows: S211, the input image is subjected to a convolution operation with a kernel of 3x3 to obtain a set of features; S212, inputting the feature into a dynamic enhancement unit DEU that works in a human eye-like mechanism to enhance the feature, so that it can better represent the brightness level of the image and enhance the brightness using statistical characteristics; S213, using a layer of 3x3 convolution and Tanh activation operation to encode the image and output it to complete the preliminary brightness improvement; The specific implementation process of the frequency domain noise suppression module is as follows: S221, transforming the image features obtained by the time domain brightness enhancement module into the frequency domain using wavelet transform DWT; S222, using a dynamic enhancement unit DEU and basic convolution to process the features; S223, gradually transforming the processed image back to a time domain image by performing an inverse wavelet transform (IDWT), and then entering the processing of a time domain detail reconstruction module; The specific implementation process of the time domain detail reconstruction module is as follows: S231, cascading the output result graphs of the time domain brightness enhancement module and the frequency domain noise suppression module, and inputting them into the time domain detail reconstruction module; S232, the input image is subjected to several convolutional layers and multiple dynamic enhancement units DEU to perform image edge enhancement, color restoration and detail information reconstruction; S233, output the final enhanced image of the entire network.

4. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 3, characterized in that: The specific process of step S211 is: the input image S∈R with image size H×W and channel number 3 H×w×3 , a convolutional layer with a kernel of 3×3 and a leaky rectified linear unit activation operation are used to extract a set of features of size [8, 64, 256, 256] containing the brightness, color, and detail information of the image.

5. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 3, characterized in that: The preliminary brightness enhancement process in step S213 can be expressed as: S l =Tanh(Rconv3(F deu (Rconv3(S)))) (5) Where S represents the input image; F deu (·) is the operation of the dynamic enhancement unit DEU; Tanh(·) is the hyperbolic tangent activation function; S l Represents the output features.

6. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 3, characterized in that: After the features enter the time domain detail reconstruction module, they are sequentially subjected to 1×1 convolution, 3×3 convolution, three dynamic enhancement units DEU, and 3×3 convolution operations, and driven by the loss function; the loss function is defined as follows: L total = l1+l2+l3 (10) l1=||S l -S normal ||1 (11) l2=||S d -S normal ||1+||ΔS d ||1 (12) l3=||S s -S normal ||1+||F lap (S s )-F lap (S normal )||1 (13) Among them, L total represents the total loss of the low-light enhancement network; l1, l2, and l3 are the losses of the temporal brightness enhancement module, the frequency domain noise suppression module, and the temporal detail reconstruction module, respectively; S l ,S d ,S s They are the output result diagrams of the time domain brightness enhancement module, the frequency domain noise suppression module, and the time domain detail reconstruction module; S normal is the label image; ||·||1 is the L1 regularization function; Δ represents the sum of the horizontal and vertical gradients of the image; F lap (·) is the Laplace transform function.

7. The low-light image enhancement method using time-frequency domain hierarchical processing according to claim 1, characterized in that: The specific process of step S3 is as follows: S31. When training the model, the input image size is adjusted to 256×256. The network is trained using the adaptive moment estimation optimizer with a running average coefficient of (0.9, 0.999) and a weight decay of 0.000008. The number of training iterations is 150, the batch size is 8, the initial learning rate is 0.0002, and the learning plan strategy is adopted. After the number of iterations reaches 60, the learning rate decreases by 10%; S32. Save the model after every 10 training iterations.