A low-light image enhancement method based on Retinex, a storage medium and a terminal

Abstract

This invention discloses a Retinex-based low-light image enhancement method, storage medium, and terminal. The method includes: enhancing the luminance channel of a first low-light image using a luminance processing module to obtain a luminance-enhanced channel; recombining the luminance-enhanced channel with other color channels to form a first enhanced image that meets luminance requirements but suffers from color degradation; processing the color degradation of the first enhanced image using a color restoration module to obtain a second enhanced image; performing supervised learning on the second enhanced image using a normal image to construct a low-light model applicable to low-light images in all scenes; and inputting the second low-light image into the pre-trained low-light model to obtain the enhanced target image. This invention applies Retinex theory to enhance brightness while preserving image texture, addressing texture degradation and color degradation in two steps, thus more effectively handling various types of degradation and improving the quality of low-light images in terms of brightness, contrast, color, and texture detail.

Description

Technical Field

[0001] This invention belongs to the field of digital image processing and computer vision technology, specifically a low-light image enhancement method, storage medium, and terminal based on Retinex. Background Technology

[0002] Low-light images often suffer from complex degradations such as low visibility, low contrast, tone shift, texture loss, and noise. Current model-based and deep learning-based low-light enhancement methods often only consider a few of these degradations or address them all together. These methods may achieve good results in one aspect of brightness, color, texture detail, or noise reduction, but they rarely achieve satisfactory results in all of them simultaneously. RRM effectively removes noise but performs poorly in improving image brightness; Retinex-Net (Retinex is a computational theory of color constancy perception, a compound word composed of retina and cortex) effectively improves image brightness but suffers from severe noise and tone shift; Zero-DCE (Zero-Reference Deep Curve Estimation) solves color shift well but has low contrast and unclear texture; RUAS improves low image contrast but often leads to local overexposure; AGLLNet (attention-guided low-light Net) performs well in color and brightness but has poor local texture recovery and often produces artifacts; URetinex-Net performs well in texture and brightness but has low saturation and the image is not vivid enough.

[0003] Therefore, there is an urgent need for a Retinex-based low-light image enhancement method that can improve the quality of low-light images in terms of brightness, contrast, color, and texture details. Summary of the Invention

[0004] The purpose of this invention is to provide a Retinex-based low-light image enhancement method, storage medium, and terminal to address the technical problem that current model-based and deep learning-based low-light enhancement methods often only consider a few types of degradation or address these complex and diverse degradations together. This results in these methods potentially achieving good results in one aspect of brightness, color, texture detail, or noise processing, but failing to achieve satisfactory results in all of these aspects simultaneously.

[0005] Therefore, the present invention provides a low-light image enhancement method based on Retinex, comprising:

[0006] Input a first low-light image, and the pre-built brightness processing module enhances the brightness channel of the first low-light image to obtain a brightness enhancement channel;

[0007] The brightness enhancement channel is recombined with other color channels to obtain a first enhanced image that meets the brightness requirements but suffers from color degradation. The other color channels are color channels other than the brightness channel.

[0008] A pre-built color restoration module processes the color degradation of the first enhanced image to obtain a second enhanced image;

[0009] Supervised learning is performed on the second enhanced image using the normal image to update the parameters of the brightness processing module and the color restoration module, so as to build a low-light model applicable to low-light images in all scenes.

[0010] The second low-light image is input into the pre-trained low-light model to obtain the enhanced target image.

[0011] In a more preferred embodiment, the input first low-light image is further enhanced by a pre-built brightness processing module to obtain a brightness enhancement channel, which includes:

[0012] An initial decomposition module is constructed to initially decompose the brightness channel of the image into R0 and L0, where R0 is a coarsely estimated reflection component of the object and L0 is a coarsely estimated ambient light component.

[0013] An expansion solution module is constructed to solve the Retinex decomposition problem on the brightness channel, and accurate R and L are obtained, where R is the reflection component of the object and L is the ambient light component.

[0014] A brightness fusion module is constructed, which integrates the obtained accurate reflection components of the object with an additional supplementary illumination map to obtain a brightness enhancement channel. The additional supplementary illumination map is an image with uniform and sufficient brightness.

[0015] In a more preferred embodiment, solving the Retinex decomposition problem on the luminance channel using the expansion solver module includes:

[0016] Retinex decomposition on the luminance channel:

[0017]

[0018] Among them, ‖VR·L‖ F It is the F-norm. For the reconstruction terms, V is the luminance channel, Ω1(R) and Ω2(L) are the regularization terms for R and L respectively, and α1 and α2 are hyperparameters.

[0019] In a more preferred embodiment, solving the Retinex decomposition problem on the luminance channel using the expansion solver module further includes:

[0020] The ambient light component L is calculated from the estimated R by supervised learning.

[0021] L = V / R

[0022] Based on the obtained ambient light component L, the Retinex decomposition on the luminance channel is performed:

[0023]

[0024] Where α is a hyperparameter and Ω(R) is the regularization term of R.

[0025] In a more preferred embodiment, solving the Retinex decomposition problem on the luminance channel using the expansion solver module further includes:

[0026] Introducing an auxiliary variable Z, we perform Retinex decomposition on the luminance channel:

[0027]

[0028] The Retinex decomposition problem is broken down into three univariate subproblems:

[0029]

[0030]

[0031]

[0032] Where Ω(Z) is the regularization term of Z, γ is the hyperparameter, and ||ZR|| F It is the F-norm. For the subproblem related to the auxiliary variable Z, For the subproblems related to variable R, For the subproblems related to variable L, the subproblems and These are all least squares problems, which can be solved by making their first derivative equal to 0.

[0033] In a more preferred embodiment, the step of solving the Retinex decomposition problem on the luminance channel using the expansion solver module then includes:

[0034] The Retinex decomposition problem on the luminance channel is solved by iteratively updating variables. The formula for each iteration is:

[0035]

[0036]

[0037]

[0038] in For a convolutional neural network, θ z It is a convolutional neural network The parameters are k, which is the number of iterations, and R. k Let R be the value of R during the k-th iteration update. k+1 Let R be the value of Z during the (k+1)th iteration update. k+1 Let Z and L be the values ​​of Z and L during the (k+1)th iteration update. k Let L be the value of L during the k-th iteration update. k+1 This is the value of L during the (k+1)th iteration update.

[0039] In a more preferred embodiment, the color restoration module includes a backbone network, a first downsampling branch network, and a second downsampling branch network; the backbone network includes 8 convolutional layers and activation functions, wherein the activation function of the first 7 convolutional layers is LeakyReLU, and the activation function of the 8th layer is ReLU.

[0040] In a more preferred embodiment, the initialization decomposition module is an initialization convolutional neural network, which includes four convolutional layers and activation functions, wherein the activation function of the first three convolutional layers is LeakyReLU, and the activation function of the last convolutional layer is ReLU.

[0041] On the other hand, the present invention provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the Retinex-based low-light image enhancement method as described above.

[0042] On the other hand, the present invention also provides a terminal, comprising: a processor and a memory, wherein the memory stores a computer-readable program executable by the processor; when the processor executes the computer-readable program, it implements the steps in the Retinex-based low-light image enhancement method as described above.

[0043] Compared with existing technologies, the features and beneficial effects of this invention are as follows: For complex degradation in low-light images, this invention adopts a step-by-step approach, breaking down complex problems into multiple relatively simple ones for solution. Considering that these degradations are not entirely independent, this invention roughly categorizes them into two main types: texture degradation and color degradation, and models and solves them sequentially. To handle texture degradation, this invention first focuses on brightness adjustment and texture preservation, proposing a Retinex-based brightness processing module (LPM) that enhances only the brightness channel of the first low-light image, resulting in a brightness enhancement channel. After processing texture degradation, this invention recombines the brightness enhancement channel with other color channels to obtain a first enhanced image that meets brightness requirements but suffers from color degradation. Then, a color restoration module is used to process the color degradation of the first enhanced image, resulting in a second enhanced image. This invention enhances brightness while preserving image texture by applying Retinex theory, and solves texture degradation and color degradation in two steps. The constructed model can more effectively handle various types of degradation compared to other methods, improving the quality of low-light images in terms of brightness, contrast, color, and texture detail. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the low-light image enhancement method based on Retinex.

[0046] Figure 2 A schematic diagram of the overall framework of the low-light model;

[0047] Figure 3 This is a schematic diagram of the SE network structure;

[0048] Figure 4 and Figure 5 This is a visual comparison chart of the method of this patent with other methods;

[0049] Figure 6 The structural schematic diagram of the terminal device provided by the present invention.

[0050] Attached image labels: 10-Processor, 11-Display, 12-Memory, 13-Communication interface, 14-Bus. Detailed Implementation

[0051] This invention provides a low-light image enhancement method, storage medium, and terminal based on Retinex. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0052] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application’s specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when the invention refers to an element as being “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0053] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0054] In today's digital age, people's lives are inseparable from a large amount of multimedia information, including images, videos, and audio. Among these, visual perception information, represented by images and videos, is the most direct and important part of the information people acquire. Digital images are widely used in security monitoring, traffic management, satellite remote sensing, military reconnaissance, and artificial intelligence fields such as facial recognition, semantic segmentation, and autonomous driving. It can be said that the vast majority of human activities rely on the application of images. In order to fully capture the information in images, people's requirements for image quality are constantly increasing. However, during image acquisition, noise is inevitably introduced due to interference from the external environment, leading to a decrease in image quality. In low-light environments, due to the lack of sufficient light, the acquired images contain large areas of darkness and severe noise, resulting in a significant decrease in image visibility and contrast. This seriously affects the effectiveness of subsequent high-level vision tasks, such as reducing the accuracy of facial recognition and license plate recognition.

[0055] To acquire high-quality images in low-light environments, the most direct approach is to adjust the parameters of the physical imaging process to improve image quality. Extending the exposure time of the imaging device allows for the capture of more light, but it is highly susceptible to blurring due to camera shake. Another method is to increase the film's sensitivity, which significantly improves image brightness but also introduces considerable noise, reducing image detail and overall quality. Using a flash or external lighting is also a common technique, but uneven light distribution can lead to unnatural colors. In short, simply relying on adjusting the physical imaging process is insufficient to obtain ideal images; therefore, developing intelligent low-light enhancement algorithms to acquire high-quality images is essential.

[0056] Based on different algorithm design philosophies, existing low-light image enhancement algorithms can be divided into three categories: distribution mapping-based methods, model optimization-based methods, and deep learning-based methods.

[0057] Distribution mapping-based methods consider the pixel value distribution of low-light images and use techniques such as gamma correction, power functions, logarithmic functions, histogram equalization, and adaptive histogram equalization to improve the pixel value distribution and thus increase image brightness. However, this technique, which only considers pixel value distribution and ignores the imaging process and the spatial distribution of pixel values, was a common approach to image processing in the early stages of artificial intelligence. It often fails to effectively distinguish the semantic information of images and suffers from color distortion, significant noise, and unclear texture.

[0058] Model optimization methods often employ Retinex theory to mathematically model images and iteratively update model parameters by designing an objective optimization function to obtain a model that effectively enhances the image, thereby achieving image enhancement. Retinex theory simulates the human visual system, assuming that an image I is composed of the product of a reflection map R and an illumination map L, i.e., I = R·L, where the reflection map R represents the scene information of the image, and the illumination map L represents the external lighting information. The difference between low-light and normal-light images lies in the distribution and intensity of external lighting; therefore, in Retinex theory, R is the same for both, but L is different. Thus, Retinex-based model optimization methods decompose the image to obtain R and L, then enhance L, and finally multiply the enhanced L by R to achieve low-light enhancement. However, these methods require manually adding prior regularization terms for Retinex decomposition and L enhancement, i.e., setting constraints and optimization directions for variables based on experience. This results in poor generalization performance of the model, often only achieving good enhancement results in specific scenarios and being difficult to apply to complex scenarios. Furthermore, a model needs to be trained for each image, which makes batch image processing very time-consuming and difficult to apply to real-time image enhancement (such as camera night scene shooting optimization).

[0059] In the era of big data, in order to overcome the shortcomings of the traditional methods mentioned above, the application of the data-driven model concept and the use of the powerful nonlinear mapping capabilities of deep learning to establish the relationship between low-light input and enhanced output by designing network structures has become a mainstream low-light image enhancement mode.

[0060] To address the problem that current model-based and deep learning-based low-light image enhancement methods often only consider a few types of degradation or address these complex and diverse degradations together, while achieving good results in one aspect such as brightness, color, texture detail, and noise reduction, it is difficult to achieve satisfactory results in all aspects simultaneously, this invention proposes a Retinex-based low-light image enhancement method, which is a novel supervised low-light image enhancement method based on deep learning.

[0061] Example 1

[0062] like Figure 1 and Figure 2 As shown, taking the HSV color space as an example, the Retinex-based low-light image enhancement method specifically includes:

[0063] S10. Input a first low-light image. The pre-built Luminance Processing Module (LPM) enhances the V channel (luminance channel) of the first low-light image to obtain the Venhanced luminance channel. Specifically, the first low-light image is first converted from the RGB color space to the HSV color space, and the LPM is used to enhance the luminance channel to obtain the Venhanced luminance channel. The first low-light image is a low-light image from the training samples.

[0064] The construction method of the luminance processing module includes building an initialization decomposition module (Init), an unfolding block, and an luminance fusion module (BIF). The initialization decomposition module initially decomposes the luminance channels of the image into R0 and L0, where R0 is the original reflection component of the object and L0 is the original ambient light illumination component. The unfolding block solves the Retinex decomposition problem on the luminance channels to obtain accurate R and L. The luminance fusion module integrates the additional, uniform, and sufficiently bright supplementary illumination map. Together with R, it is fed into the luminance blending module BIF, and the enhanced luminance enhancement channel Venhanced is output.

[0065] In this embodiment, the initialization decomposition module is the initialization convolutional neural network (Init). The initialization convolutional neural network (Init) consists of four convolutional layers and activation functions. The first three convolutional layers use LeakyReLU activation functions, and the last convolutional layer uses ReLU activation functions. It should be noted that the initialization decomposition module can also be implemented using other traditional initialization methods, such as:

[0066] The Retinex decomposition problem on the luminance channel can generally be described as follows:

[0067]

[0068] Among them, "VR·L" F It is the F-norm. For the reconstruction terms, V is the luminance channel, R is the reflection component of the object, L is the ambient light component, Ω1(R) and Ω2(L) are the regularization terms of R and L respectively, and α1 and α2 are hyperparameters.

[0069] Since the normal image (Inormal) contains sufficient scene information, and the reflectance image (R) represents the scene information in the image, this invention estimates R through supervised learning during Retinex decomposition of the luminance channel, and then uses formula (1) to obtain L. Formula (1) is:

[0070] L=V / R (1)

[0071] Therefore, this invention discards the regularization term on L to reduce computation and shorten the runtime by nearly half. The Retinex decomposition problem on the luminance channel is then:

[0072]

[0073] To more easily solve this optimization problem, this invention introduces an auxiliary variable Z during the Retinex decomposition of the luminance channel. The Retinex decomposition problem of the luminance channel can then be transformed into:

[0074]

[0075] In formula (4), γ is a hyperparameter, Ω(Z) is the regularization term of Z, and ||ZR|| F It is the F-norm;

[0076] Formula (4) can then be broken down into three single-variable subproblems:

[0077]

[0078]

[0079]

[0080] in For the subproblem related to the auxiliary variable Z, For the subproblems related to variable R, This is a subproblem related to variable L. In traditional methods, Ω(Z) is assumed to be designed; however, this invention uses a convolutional neural network. Ω(Z) is learned from the data, making it applicable to different scenarios. Subproblems and These are all least squares problems, which can be solved by making their first derivative equal to 0.

[0081] This invention uses an iterative variable update method to solve the Retinex decomposition problem on the luminance channel. The formula for each iteration update is:

[0082]

[0083]

[0084]

[0085] in For a convolutional neural network, θ z It is a convolutional neural network The parameters are k, which is the number of iterations, and R. k Let R be the value of R during the k-th iteration update. k+1 Let R be the value of Z during the (k+1)th iteration update. k+1 Let Z and L be the values ​​of Z and L during the (k+1)th iteration update. k Let L be the value of L during the k-th iteration update. k+1 Let L be the value at the (k+1)th iteration update. In a specific embodiment, k is set to 3, which balances runtime and model performance. More iterations result in more accurate estimates of R and L, but also increase time overhead. Those skilled in the art can set k to other specific values ​​based on experimental conditions.

[0086] After passing through the initialization decomposition module and the expansion solution module, the brightness channel of the first low-light image is decomposed into R and L. Since L is only related to the illumination at the time of image acquisition, the L of the first low-light image and the normal-light image can be independent and unrelated. Therefore, unlike other existing methods, this invention does not map L from low-light to normal-light, but instead supplements it with a new illumination map that is uniformly bright and sufficiently bright. Together with R, it is fed into the luminance blending module BIF, which outputs an enhanced luminance enhancement channel, Venhanced. This process can be described as follows:

[0087]

[0088] Where RK is R in the last iteration of the expansion solver module, and θ IF These are the parameters for the BIF (Brightness Identifier) ​​of the neural network. This is to ensure the availability of additional supplementary lighting maps. It is uniform. This invention first calculates the mean ω of the illumination map Lnormal of the normal illumination image, and then expands ω to the same size as RK to obtain an additional supplementary new illumination map. The formula for calculating ω is:

[0089]

[0090] The Brightness Blending (BIF) module consists of 5 convolutional layers plus activation functions, with the first 4 layers using LeakyReLU and the last layer using ReLU.

[0091] S20. Recombine the luminance enhancement channel (Venhanced) with other color channels to obtain the first enhanced image I that meets the luminance requirements but suffers from color degradation. obe Other color channels refer to color channels other than the luminance channel. Specifically, taking the HSV color space as an example, the luminance enhancement channel Venhanced is recombine with the H and V color channels to obtain a first enhanced image I that meets the luminance requirements but suffers from color degradation. obe .

[0092] S30. The pre-built color restoration module CRM processes the color degradation of the first enhanced image to obtain the second enhanced image Ienhanced.

[0093] CRM consists of a color restoration network CRNet, whose color restoration module includes a backbone network, a first downsampling branch network, and a second downsampling branch network. The backbone network comprises eight convolutional layers and activation functions. The first seven convolutional layers use LeakyReLU activation, and the eighth layer uses ReLU. The feature maps from the first convolutional layer of the backbone network are stacked and concatenated with the feature maps from the seventh convolutional layer before participating in the next convolution. Similarly, the second convolutional layer of the backbone network is stacked and concatenated with the feature maps from the sixth convolutional layer before participating in the next convolution. The third convolutional layer of the backbone network is also stacked and concatenated with the feature maps from the fifth convolutional layer before participating in the next convolution. Similarly, the first convolutional layer of the downsampling branch network is stacked and concatenated with the feature maps from the third convolutional layer before participating in the next convolution.

[0094] After the backbone network undergoes two convolutions, the resulting feature map is downsampled and enters the first downsampling branch network. After four convolutional layers and activation functions, the first downsampling branch network upsamples the resulting feature map and concatenates it with the feature map obtained from the sixth convolution of the backbone network, which then participates in the next convolution.

[0095] After one convolution, the downsampling first branch network performs downsampling and then enters the downsampling second branch network. After two convolutional layers and activation functions, the downsampling second branch network performs upsampling and concatenates the feature map obtained from the third convolution of the downsampling first branch network with the feature map to participate in the next convolution.

[0096] Furthermore, the feature maps concatenated after downsampling in both the first and second downsampling branches use the SE module for feature fusion. The structure of the SE module is as follows: Figure 3 As shown in Table 1, the use of a downsampling first branch network and a downsampling second branch network effectively expands the receptive field of the feature map, which is beneficial for restoring image details. Introducing the SE module effectively fuses image features, which is beneficial for overall image color restoration. Detailed information about the CRNet network is shown in Table 1.

[0097] Table 1. Detailed information about the CRNet network

[0098]

[0099]

[0100] S40. Supervised learning is performed on the second enhanced image using the normal image to update the parameters of the brightness processing module and the color restoration module, so as to build a low-light model applicable to low-light images in all scenes.

[0101] The loss function for training the initialization decomposition module is as follows:

[0102]

[0103] Where En is the identity matrix, and the second constraint term This is to prevent L0 from being all 0, and T is the matrix transpose operation.

[0104] The expansion and solution module and the brightness fusion module are trained together, and their loss function is:

[0105]

[0106] Where φ(R) K ) represents the features extracted from RK by the VGG19 neural network feature extractor, φ(V enhanced ) represents the features extracted from Venhanced by the VGG19 neural network feature extractor, and SSIM(Venhanced V normal ) represents the structural similarity between Venhanced and Vnormal, γ is a hyperparameter, ZK is Z in the last iteration of the expansion solver, and Vnormal is the normal brightness channel.

[0107] The loss function during the training of the color restoration module is:

[0108]

[0109] In formula (15), Inormal represents a normal image. SSIM(I) is the gradient operator, which includes both horizontal and vertical directions. enhanced ,I normal ) represents the structural similarity between Ienhanced and Inormal.

[0110] S50. Input the second low-light image into the pre-trained low-light model to obtain the enhanced target image. The second low-light image is the image in the input sample.

[0111] By applying Retinex theory to enhance brightness while preserving image texture, and addressing texture degradation and color degradation in two steps, the model of this invention can more effectively handle various types of degradation compared to other methods, improving the quality of low-light images in terms of brightness, contrast, color, and texture detail.

[0112] Figure 4 and Figure 5 This invention compares the enhancement method of the present invention with other methods. Other methods may perform well in one of the aspects of brightness, texture, noise suppression, and color, while the method of the present invention can not only effectively improve brightness, contrast, and noise suppression while maintaining texture clarity, but also achieve more natural color restoration.

[0113] It should be noted that the brightness adjustment module in this invention can be used not only for the brightness channels in the HSV color space, but also for the brightness channels in other color spaces such as Lab, YUV, and YCbCr.

[0114] Example 2

[0115] Based on the above-described Retinex-based low-light image enhancement method, this invention provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the above-described Retinex-based low-light image enhancement method.

[0116] Example 3

[0117] Based on the aforementioned Retinex-based low-light image enhancement method, this invention also provides a terminal, such as... Figure 6 As shown, the system includes at least a processor 10, a display screen 11, and a memory 12, and may also include a communication interface 13 and a bus 14. The processor 10, display screen 11, memory 12, and communication interface 13 can communicate with each other via the bus 14. The display screen 11 is configured to display a preset user guide interface in the initial setup mode. The communication interface 13 can transmit information. The processor 10 can call logical instructions in the memory 12 to execute the methods described in the above embodiments.

[0118] Furthermore, the logic instructions in the aforementioned memory 12 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0119] The memory 12, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 10 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 12, thereby implementing the methods in the above embodiments.

[0120] The memory 12 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 12 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks; these may also be transient storage media.

[0121] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0122] In summary, this invention addresses complex degradation in low-light images by breaking down the complex problem into several relatively simpler ones. Considering that these degradations are not entirely independent, this invention roughly categorizes them into two main types: texture degradation and color degradation, and models and addresses them sequentially. To handle texture degradation, this invention first focuses on brightness adjustment and texture preservation, proposing a Retinex-based brightness processing module (LPM) that enhances only the brightness channel of the first low-light image, resulting in a brightness enhancement channel. After processing texture degradation, this invention recombines the brightness enhancement channel with other color channels to obtain a first enhanced image that meets brightness requirements but suffers from color degradation. Then, a color restoration module is used to process the color degradation of the first enhanced image, resulting in a second enhanced image. By applying Retinex theory to enhance brightness while preserving image texture, and addressing texture and color degradation in two steps, this invention's model more effectively handles various types of degradation compared to other methods, improving the quality of low-light images in terms of brightness, contrast, color, and texture detail.

[0123] This invention replaces the traditional manual design of the constraint term Ω(R) on R in the classic Retinex decomposition problem with adaptive fitting using a convolutional neural network. After estimating R, L can be obtained using L = V / R. Therefore, this invention discards the regularization term on L to reduce the amount of computation and shorten the running time by nearly half.

[0124] Traditional model-based optimization methods use manually designed regularization terms for R and L, which are difficult to apply to complex scenarios. This invention combines deep learning with model-based optimization, employing an unfolded solution method (iterative variable updates) to solve the Retinex decomposition problem. This invention uses a convolutional neural network to learn an adaptive regularization term from the data, which significantly improves the model's generalization performance.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A Retinex-based low-light image enhancement method, characterized in that include: Input a first low-light image, and the pre-built brightness processing module enhances the brightness channel of the first low-light image to obtain a brightness enhancement channel; The brightness enhancement channel is recombined with other color channels to obtain a first enhanced image that meets the brightness requirements but suffers from color degradation. The other color channels are color channels other than the brightness channel. A pre-built color restoration module processes the color degradation of the first enhanced image to obtain a second enhanced image; Supervised learning is performed on the second enhanced image using the normal image to update the parameters of the brightness processing module and the color restoration module, so as to build a low-light model applicable to low-light images in all scenes. The second low-light image is input into the pre-trained low-light model to obtain the enhanced target image; The input is a first low-light image. A pre-built brightness processing module enhances the brightness channel of the first low-light image to obtain a brightness enhancement channel. This process includes: An initial decomposition module is constructed to initially decompose the brightness channel of the image into R0 and L0, where R0 is a coarsely estimated reflection component of the object and L0 is a coarsely estimated ambient light component. An expansion solution module is constructed to solve the Retinex decomposition problem on the brightness channel, and accurate R and L are obtained, where R is the reflection component of the object and L is the ambient light component. A brightness fusion module is constructed, which integrates the obtained accurate reflection components of the object with an additional supplementary illumination map to obtain a brightness enhancement channel. The additional supplementary illumination map is an image with uniform and sufficient brightness.

2. The Retinex-based low-light image enhancement method of claim 1, wherein, The method of solving the Retinex decomposition problem on the luminance channel using the expansion solution module further includes: Retinex decomposition on the luminance channel: in, It is the F-norm. For the reconstruction item, V represents the luminance channel. and These are the regularization terms for R and L, respectively. and This is a hyperparameter.

3. The Retinex-based low-light image enhancement method of claim 2, wherein, The method of solving the Retinex decomposition problem on the luminance channel using the expansion solution module further includes: The ambient light component L is calculated from the estimated R by supervised learning. Based on the obtained ambient light component L, the Retinex decomposition on the luminance channel is performed: wherein is a hyperparameter, is a regularizer for R.

4. The low-light image enhancement method based on Retinex according to claim 3, characterized in that, The method of solving the Retinex decomposition problem on the luminance channel using the expansion solution module further includes: Introducing an auxiliary variable Z, we decompose the Retinex on the luminance channel: The Retinex decomposition problem is broken down into three univariate subproblems: in For Z, For hyperparameters, It is the F-norm. For the subproblem related to the auxiliary variable Z, For the subproblems related to variable R, For the subproblems related to variable L, the subproblems and These are all least squares problems, which can be solved by making their first derivative equal to 0.

5. The Retinex-based low-light image enhancement method of claim 1, wherein, The process of solving the Retinex decomposition problem on the luminance channel using the expansion solution module then includes: The Retinex decomposition problem on the luminance channel is solved by iteratively updating variables. The formula for each iteration is: in It is a convolutional neural network. It is a convolutional neural network The parameter k is the number of iterations. Let R be the value at the k-th iteration update. The value of R is the value updated in the (k+1)th iteration. Let Z be the value updated in the (k+1)th iteration. Let L be the value at the time of the k-th iteration update. This is the value of L during the (k+1)th iteration update.

6. The Retinex-based low-light image enhancement method of claim 1, wherein, The color restoration module includes a backbone network, a first downsampling branch network, and a second downsampling branch network; the backbone network includes 8 convolutional layers and activation functions, wherein the activation function of the first 7 convolutional layers is LeakyReLU, and the activation function of the 8th layer is ReLU.

7. The Retinex-based low-light image enhancement method of claim 1, wherein: The initialization decomposition module initializes a convolutional neural network, which includes four convolutional layers and activation functions. The activation function for the first three convolutional layers is LeakyReLU, and the activation function for the last convolutional layer is ReLU.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the Retinex-based low-light image enhancement method as described in any one of claims 1 to 7.

9. A terminal, characterized by include: A processor and a memory, wherein the memory stores a computer-readable program that can be executed by the processor; when the processor executes the computer-readable program, it implements the steps of the Retinex-based low-light image enhancement method as described in any one of claims 1 to 7.

Images