Low-illumination image enhancement method based on improved Retinex-Net and intelligent wearing system

Abstract

The invention provides a low-illumination image enhancement method based on improved Retinex-Net and an intelligent wearing system, and the method comprises the following steps: collecting a data set with a low-illumination label, and dividing the data set into a training set and a test set; a low-illumination image enhancement model based on improved Retinex-Net is designed and trained and verified, a trained low-illumination image enhancement model is obtained, and the low-illumination image enhancement model is composed of a decomposition network, an adjustment network and a reconstruction network. A sub-network formed by an identical mapping and residual shrinkage construction unit is added into a reflection path of the decomposition network; and enhancing the image by using the trained low-illumination image enhancement model to obtain an enhanced image. According to the low-illumination image enhancement method and the intelligent wearing system provided by the invention, noise generated in an image decomposition process can be effectively suppressed, the definition and quality of the image are improved, and low-illumination image enhancement is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement and intelligent wearable systems, and in particular to a low-illumination image enhancement method and an intelligent wearable system based on an improved Retinex-Net. Background Art

[0002] With advances in modern transportation technology and a stringent safety management system, accident rates have significantly decreased. Despite overall safety improvements, drivers and pilots still face specific risks. Weather conditions are a significant challenge to operational safety, particularly in low-visibility environments. Visibility refers to the maximum horizontal distance at which a person with normal vision can identify an object under current weather conditions, directly impacting vehicle takeoff, landing, and operational safety. Image enhancement, a key technology for improving image readability, can effectively enhance the operator's field of view in low-light conditions. Traditionally, pilots rely on instrument flight rules (IFR), while drivers reduce speed and activate fog lights. However, in extremely low visibility, safety risks such as misjudgment and delayed response still exist. Therefore, intelligent visibility enhancement technology has become an essential industry need. The rapid development of artificial intelligence (AI) has provided new opportunities. Its high-precision processing and all-weather operation make it an effective solution to operational challenges in low-visibility environments. This technology has achieved significant results in fields such as intelligent transportation, autonomous driving, and security surveillance, making the development of intelligent vision enhancement systems applicable to multiple scenarios increasingly urgent.

[0003] However, while AI is currently being applied in transportation, dedicated solutions for drivers and pilots remain insufficient. Traditional approaches to dealing with low visibility are limited by inefficiency and poor adaptability, and existing augmentation technologies struggle to meet the real-time demands of complex environments.

[0004] Therefore, proposing a new low-light image enhancement method and intelligent wearable system to assist pilots and drivers in driving safely has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a low-light image enhancement method and an intelligent wearable system based on an improved Retinex-Net to solve the problems existing in traditional low-light image enhancement methods.

[0006] On one hand, the present invention provides a low-light image enhancement method based on an improved Retinex-Net, comprising:

[0007] S1: Collect a dataset with low-light labels and divide it into a training set and a test set, wherein the dataset includes images of different scenes and different lighting conditions;

[0008] S2: Designing a low-light image enhancement model based on the improved Retinex-Net and using the training set and the test set to train and verify the low-light image enhancement model to obtain a trained low-light image enhancement model;

[0009] The low-light image enhancement model consists of three parts: a decomposition network, an adjustment network, and a reconstruction network. A subnetwork formed by identity mapping and residual shrinkage building blocks is added to the reflection path of the decomposition network. The training process of the model is as follows:

[0010] S21: Using the decomposition network to transform the input low-light image S low With the normal image S normal Corresponding decomposition into reflection component (R low ,R normal ) and the illumination component (I low ,I normal ), and the reflection component R of the low illumination image low Optimize and obtain the optimized reflection component R of the low-light image low1 ;

[0011] S22: The reflection component R of the optimized low-light image is constructed using a sub-network formed by the identity mapping and residual shrinkage construction unit. low1 Perform noise suppression to obtain the reflection component R of the denoised low-light image denoised ;

[0012] S23: Using the adjustment network to adjust the illumination component I of the low-light image low Correction is performed to obtain the corrected illumination component I corrected ;

[0013] S24: Using the reconstruction network to reconstruct the denoised reflection component R denoised and the corrected illumination component I corrected Element-by-element multiplication to restore the enhanced image corresponding to the low-light image;

[0014] S25: Calculating the error between the enhanced image and the normal image using a loss function and performing back propagation to update the parameters of the low-illumination image enhancement model until the training of the low-illumination image enhancement model is completed;

[0015] S3: Using the trained low-light image enhancement model, the low-light image collected in real time is enhanced to obtain an enhanced image.

[0016] Preferably, in S1, the training set uses the LOL training set and BrightingTrain, and the test set uses the DICM dataset and the LOL evaluation set.

[0017] Further preferably, in S21, the decomposition network is a dual-branch structure, the dual branches include an illumination branch and a reflection branch, the illumination branch uses 7 layers of 7×7 convolutional layers based on the Sigmoid activation function to extract the illumination component (I low ,I normal ); the reflection branch calculates the reflection component (R low ,R normal ), and optimize the reflection component R of the low-light image using 3 layers of 3×3 convolutional layers based on the ReLU activation function low , get the optimized reflection component R low1 .

[0018] Further preferably, in S21, 7 layers of 7×7 convolution layers based on Sigmoid activation function are used to extract the illumination component (I low ,I normal ) is as follows:

[0019] I low =Sigmoid(Conv 7×7 (F shallow ));

[0020] I normal =Sigmoid(Conv 7×7 (F shallow ));

[0021] Where, F shallow The low-light image S is input using three 3×3 shallow convolutions low With the normal image S normal The calculation formula for the shared basic features obtained after feature extraction is as follows:

[0022] F shallow =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (S)))));

[0023] Where, the ReLU function is defined as ReLU(x)=max(0,x); S is S low and S normal .

[0024] More preferably, in S21, the reflection component (R low ,R normal ) is as follows:

[0025] R low =S low -I low;

[0026] R normal =S normal -I normal ;

[0027] Utilize 3 layers of 3×3 convolutional layers based on ReLU activation function to optimize the reflection component R of the low-light image low , get the optimized reflection component R low1 The formula is as follows:

[0028] R low1 =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (R low ))))))).

[0029] Further preferably, the specific steps of S22 are as follows:

[0030] S221: Calculate the identity mapping path Identity using the following formula:

[0031] Identity=Conv 1×1 (R low1 );

[0032] S222: Calculate the residual shrinkage path. The specific steps are as follows:

[0033] First, calculate the residual F(x) using the following formula:

[0034] F(x)=Conv 3×3 (R low1 );

[0035] Secondly, the adaptive threshold is generated using the following formula

[0036]

[0037] Where σ is the Sigmoid activation function; W1 is the weight matrix of the first fully connected layer, with a size of C×C / r, where C is R low1 The number of channels, r is the reduction ratio; W2 is the weight matrix of the second fully connected layer, with a size of C / r×C; GAP represents global average pooling;

[0038] Then, use the following formula to perform soft threshold denoising to obtain the denoised residual F denoised (x):

[0039]

[0040] In the formula, sign is ±1;

[0041] Finally, the reflection component R of the denoised low-light image is calculated using the following formula: denoised :

[0042] R denoised =ReLU(Identity+F denoised (x)).

[0043] Further preferably, the specific steps of S23 are as follows:

[0044] S231: Extracting the illumination component I of the low illumination image by downsampling with an encoder low The encoder consists of a four-level downsampling module, each of which contains two 3×3 convolutional layers and a 2×2 maximum pooling layer connected to the convolutional layer. Each convolutional layer uses ReLU activation, and the number of channels in the four levels is gradually doubled from 64 to 512. The encoder formula is as follows:

[0045] E l =MaxPool(ReLU(W l,2 *ReLU(W l,1 *E l-1 )));

[0046] Where, E l Represents the output of the encoder level l; MaxPool represents the 2×2 maximum pooling operation for downsampling; W l,1 With W l,2 is the convolution weight corresponding to the two 3×3 convolutions in the lth level; E0=I low , I low is the input illumination component;

[0047] S232: The decoder symmetrically expands the features extracted by the encoder. The decoder adopts a four-level upsampling structure. Each level first performs a double upsampling through 2×2 transposed convolution to restore the spatial resolution. Then, the channel is spliced ​​with the high-resolution features of the corresponding level of the encoder. Finally, two 3×3 convolutional layers are used to fuse the cross-scale information. Each convolutional layer uses ReLU activation. The number of channels in the four levels decreases from 512 to 64. The decoder formula is as follows:

[0048] D 4-l =Conv 3×3 (ReLU(Conv 3×3 (Concat(UpSample(D 4-l+1 ),E l ))));

[0049] Where D 4-lis the output of the (4-l)th level of the encoder, Concat represents the concatenation of the upsampling result and the encoder feature along the channel dimension, and UpSample is the upsampling operation;

[0050] S233: Calculate the illumination adjustment ΔI using residual learning. The formula is as follows:

[0051] ΔI=Tanh(W final *D1);

[0052] Where Tanh is the activation function, W final Represents the weight of the convolutional layer in residual learning;

[0053] S234: The illumination component I of the low illumination image is converted into low Added to the illumination adjustment value ΔI, the corrected illumination component I is obtained. corrected , the calculation formula is as follows:

[0054] I corrected =I low +ΔI.

[0055] Further preferably, in S25, the formula of the loss function L is as follows:

[0056] L=L rec +αL rs +βL is +γL mc ;

[0057] Among them, L rs =||R1-R h ||1;

[0058]

[0059] L mc =||M*exp(-c*M)||1;

[0060] L rec =||I h -R h *L h ||1+||I1-R1*L1||1;

[0061] Where: α, β and γ represent the weight parameters of the loss function respectively; R1 is the reflectivity of the low-light image, R h is the reflectivity of the normal image, ||·||1 is the Ll norm; is the illumination gradient of the illumination component of the normal image, is the illumination gradient of the normal image, is the illumination gradient of the illumination component of the low-light image, is the illumination gradient of the low-light image; M is and The sum of , c is the parameter that controls the shape of the function.

[0062] More preferably, in the loss function, α=0.01, β=0.15, and γ=0.21.

[0063] The present invention also provides an intelligent wearable system for executing the above-mentioned low-illumination image enhancement method based on the improved Retinex-Net under low-illumination conditions.

[0064] The low-light image enhancement method and intelligent wearable system based on the improved Retinex-Net provided by the present invention utilize a low-light image enhancement model based on Retinex-Net. A subnetwork formed by identity mapping and residual shrinkage construction units effectively decomposes low-light images and normal images into two parts: illumination component and reflection component. This can effectively suppress the noise generated during the image decomposition process, improve the clarity and quality of the image, and thus achieve low-light image enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flowchart of the low-light image enhancement method based on the improved Retinex-Net provided by the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below with reference to specific embodiments, but the present invention is not limited thereto.

[0067] Despite the application of artificial intelligence in the transportation sector, solutions specifically tailored to drivers and pilots remain insufficient. Traditional low-visibility response methods are inefficient and poorly adaptable, and existing technologies struggle to meet real-time requirements.

[0068] In response to the above problems, Figure 1 As shown, the present invention provides a low-light image enhancement method based on an improved Retinex-Net, comprising the following steps:

[0069] S1: Collect a dataset with low-light labels and divide it into a training set and a test set, wherein the dataset includes images of different scenes and different lighting conditions;

[0070] The training set uses the LOL training set and BrightingTrain, and the test set uses the DICM dataset and the LOL evaluation set.

[0071] In order to make the results more accurate, a large amount of training data is required in the learning stage. In order to solve the problem of the limited number of images in the dataset, a data enhancement step is also included. Specifically: data enhancement is performed by performing a series of random flipping, shrinking, and translation on the images in the dataset.

[0072] S2: Designing a low-light image enhancement model based on the improved Retinex-Net and using the training set and the test set to train and verify the low-light image enhancement model to obtain a trained low-light image enhancement model;

[0073] like Figure 1 As shown in the figure, the low-light image enhancement model consists of three parts: a decomposition network, an adjustment network, and a reconstruction network. A subnetwork formed by an identity mapping and a residual shrinkage construction unit is added to the reflection path of the decomposition network. The training process of the model is as follows:

[0074] S21: Using the decomposition network to transform the input low-light image S low With the normal image S normal Corresponding decomposition into reflection component (R low ,R normal ) and the illumination component (I low ,I normal ), and the reflection component R of the low illumination image low Optimize and obtain the optimized reflection component R of the low-light image low1 ;

[0075] The decomposition network is a dual-branch structure, which includes an illumination branch and a reflection branch. The illumination branch uses 7 layers of 7×7 convolutional layers based on the Sigmoid activation function to extract the illumination component (I low ,I normal ); the reflection branch is obtained by residual calculation (R low ,R normal ), and optimize the reflection component R of the low-light image using 3 layers of 3×3 convolutional layers based on the ReLU activation function low , get the optimized reflection component R low1 ;

[0076] Among them, 7 layers of 7×7 convolution layers based on Sigmoid activation function are used to extract the illumination component (I low ,I normal ) is as follows:

[0077] I low =Sigmoid(Conv 7×7 (F shallow ));

[0078] I normal =Sigmoid(Conv 7×7 (F shallow ));

[0079] Where, F shallow The low-light image S is input using three 3×3 shallow convolutions low With the normal image S normal The calculation formula for the shared basic features obtained after feature extraction is as follows:

[0080] F shallow =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (S)))));

[0081] Where, the ReLU function is defined as ReLU(x)=max(0,x), which is used to ensure that the result is non-negative; S is S low and S normal ;

[0082] The Sigmoid function is used to convert the illumination component (I low ,I normal ) is constrained to [0,1];

[0083] Among them, the reflection component (R low ,R normal ) is as follows:

[0084] R low =S low -I low ;

[0085] R normal =S normal -I normal ;

[0086] Optimize the reflection component R of the low-light image using 3 layers of 3×3 convolutional layers based on ReLU activation function low , get the optimized reflection component R of the low-light image low1 The formula is as follows:

[0087] R low1 =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (R low ))))));

[0088] Where R low1is the reflection component of the optimized low-light image;

[0089] The reflection component of the normal image is not further processed to prevent over-processing of the normal image and the introduction of artifacts or detail distortion;

[0090] S22: The reflection component R of the optimized low-light image is constructed using a sub-network formed by the identity mapping and residual shrinkage construction unit. low1 Perform noise suppression to obtain the reflection component R of the denoised low-light image denoised ;

[0091] The specific steps are as follows:

[0092] S221: Calculate the identity mapping path Identity using the following formula:

[0093] Identity=Conv 1×1 (R low1 );

[0094] S222: Calculate the residual shrinkage path. The specific steps are as follows:

[0095] First, calculate the residual F(x) using the following formula:

[0096] F(x)=Conv 3×3 (R low1 );

[0097] Secondly, the adaptive threshold is generated using the following formula

[0098]

[0099] Where σ is the Sigmoid activation function; W1 is the weight matrix of the first fully connected layer, with a size of C×C / r, where C is R low1 The number of channels, r is the reduction ratio; W2 is the weight matrix of the second fully connected layer, with a size of C / r×C; GAP represents global average pooling;

[0100] Then, use the following formula to perform soft threshold denoising to obtain the denoised residual F denoised (x):

[0101]

[0102] In the formula, sign is the sign ±1;

[0103] Finally, the reflection component R of the denoised low-light image is calculated using the following formula: denoised :

[0104] R denoised=ReLU(Identity+F denoised (x));

[0105] S23: Using the adjustment network to adjust the illumination component I of the low-light image low Correction is performed to obtain the corrected illumination component I corrected ;

[0106] S231: Extracting the illumination component I of the low illumination image by downsampling with an encoder low The encoder consists of a four-level downsampling module, each of which contains two 3×3 convolutional layers and a 2×2 maximum pooling layer connected to the convolutional layer. Each convolutional layer uses ReLU activation, and the number of channels in the four levels is gradually doubled from 64 to 512. The encoder formula is as follows:

[0107] E l =MaxPool(ReLU(W l,2 *ReLU(W l,1 *E l-1 )));

[0108] Where, E l Represents the output of the encoder at level l; MaxPool represents the 2×2 maximum pooling operation, which is used for downsampling and reducing the data dimension; W l,1 With W l,2 is the convolution weight corresponding to the two 3×3 convolutions in the first level, which is used to perform convolution operation on the input to extract features; E0=I low , I low is the input illumination component;

[0109] S232: The decoder symmetrically expands the features extracted by the encoder. The decoder adopts a four-level upsampling structure. Each level first performs a double upsampling through 2×2 transposed convolution to restore the spatial resolution. Then, the channel is spliced ​​with the high-resolution features of the corresponding level of the encoder. Finally, two 3×3 convolution layers are used to fuse the cross-scale information. The number of channels in the four levels is gradually reduced from 512 to 64. The decoder formula is as follows:

[0110] D 4-l =Conv 3×3 (ReLU(Conv 3×3 (Concat(UpSample(D 4-l+1 ),E l ))));

[0111] Where D 4-l is the output of the (4-l)th level of the encoder, Concat is the concatenation of the upsampling result and the encoder feature along the channel dimension, and UpSample is the upsampling operation;

[0112] S233: Calculate the illumination adjustment ΔI using residual learning:

[0113] ΔI=Tanh(W final *D1);

[0114] Where Tanh is the activation function, Used to constrain the amplitude of the illumination adjustment ΔI to prevent over-adjustment, W final Represents the weights of the convolutional layer in residual learning.

[0115] S234: The illumination component I of the low illumination image is converted into low Added to the illumination adjustment value ΔI, the corrected illumination component I is obtained. corrected , the calculation formula is as follows:

[0116] I corrected =I low +ΔI;

[0117] S24: Using the reconstruction network to reconstruct the denoised reflection component R denoised and the corrected illumination component I corrected Multiply element by element to restore the enhanced image S′ corresponding to the low-light image low , the formula is:

[0118]

[0119] S25: Calculate the error between the enhanced image and the normal image using the loss function and perform back propagation to update the parameters of the low-illumination image enhancement model until the training of the low-illumination image enhancement model is completed;

[0120] The formula of the loss function L is as follows:

[0121] L=L rec +αL rs +βL is +γL mc ;

[0122] Among them, L rs =||R1-R h ||1;

[0123]

[0124] L mc =||M*exp(-c*M)||1;

[0125] L rec =||I h -R h *L h||1+||I1-R1*L1||1;

[0126] Where: α, β and γ represent the weight parameters of the loss function respectively; R1 is the reflectivity of the low-light image, R h is the reflectivity of the normal image, ||·||1 is the Ll norm; is the illumination gradient of the illumination component of the normal image, is the illumination gradient of the normal image, is the illumination gradient of the illumination component of the low-light image, is the illumination gradient of the low-light image; M is and The sum of , c is the parameter that controls the shape of the function;

[0127] L rec As a loss function reflecting the decomposition accuracy, the decomposed illumination component and reflection component need to follow the Retinex theory, that is, they can be restored to the original image;

[0128] L rs It is the reflectance similarity loss. According to the previous theory, the reflectance of image decomposition under different illumination should be the same;

[0129] L is For the illumination smoothness loss, the loss is calculated for the gradients in the low light and normal directions of the illumination component;

[0130] L mc For mutual consistency loss, when the illumination gradient between low illumination and normal illumination is larger or smaller, it indicates that the illumination at this time is in balance on the surface or edge of the object. Penalty is only imposed when there is a difference between the gradients and the difference is not large.

[0131] Preferably, α=0.01, β=0.15, and γ=0.21.

[0132] S3: Use the trained low-light image enhancement model to enhance the low-light image collected in real time to obtain an enhanced image.

[0133] The low-light image enhancement method based on the improved Retinex-Net provided by the present invention addresses the current shortcomings of low-light image enhancement in generating noise. By utilizing a low-light image enhancement model based on Retinex-Net, a sub-network formed by identity mapping and residual shrinkage construction units is effectively used to decompose low-light images and normal images into two parts: illumination component and reflection component. This can effectively suppress the noise generated during the image decomposition process, improve the clarity and quality of the image, and thus achieve low-light image enhancement. Specifically, Retinex-Net has high noise in low-light image enhancement processing, which will cause loss of image details. Therefore, by introducing a sub-network formed by identity mapping and residual shrinkage construction units into the decomposition network, the noise generated during the decomposition process can be removed, thereby improving the model's ability to enhance low-light images.

[0134] The present invention also provides an intelligent wearable system for executing the above-mentioned low-illumination image enhancement method based on the improved Retinex-Net under low-illumination conditions.

[0135] The present invention applies the above-mentioned low-light image enhancement method based on the improved Retinex-Net to an intelligent wearable system, which can assist drivers or pilots in driving, and has multiple advantages: First, through the improved identity mapping and residual shrinkage construction unit, the noise generated during the image enhancement process can be effectively suppressed, preventing the noise from affecting the image quality, thereby ensuring that the visual effect after image enhancement is more natural and accurate; Second, the improved Retinex-Net can effectively enhance the details in low-light images, providing clearer visual information, helping drivers and pilots to see the surrounding environment more clearly at night or in dim environments, and improving driving or flying safety; In addition, the method can effectively enhance images in different low-light environments, whether in urban night scenes, rural roads or inclement weather, and can provide stable and reliable image enhancement effects, thereby enhancing the visual perception ability of drivers or pilots.

[0136] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0137] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0138] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0140] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0141] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A low-light image enhancement method based on improved Retinex-Net, characterized in that: include: S1: Collect a dataset with low-light labels and divide it into a training set and a test set, wherein the dataset includes images of different scenes and different lighting conditions; S2: Designing a low-light image enhancement model based on the improved Retinex-Net and using the training set and the test set to train and verify the low-light image enhancement model to obtain a trained low-light image enhancement model; The low-light image enhancement model consists of three parts: a decomposition network, an adjustment network, and a reconstruction network. A subnetwork formed by identity mapping and residual shrinkage building blocks is added to the reflection path of the decomposition network. The training process of the model is as follows: S21: Using the decomposition network to transform the input low-light image S low With the normal image S normal Corresponding decomposition into reflection component (R low ,R normal ) and the illumination component (I low ,I normal ), and the reflection component R of the low illumination image low Optimize and obtain the optimized reflection component R of the low-light image low1 ; S22: The reflection component R of the optimized low-light image is constructed using a sub-network formed by the identity mapping and residual shrinkage construction unit. low1 Perform noise suppression to obtain the reflection component R of the denoised low-light image denoised ; S23: Using the adjustment network to adjust the illumination component I of the low-light image low Correction is performed to obtain the corrected illumination component I corrected ; S24: Using the reconstruction network to reconstruct the denoised reflection component R denoised and the corrected illumination component I corrected Element-by-element multiplication to restore the enhanced image corresponding to the low-light image; S25: Calculating the error between the enhanced image and the normal image using a loss function and performing back propagation to update the parameters of the low-illumination image enhancement model until the training of the low-illumination image enhancement model is completed; S3: Using the trained low-light image enhancement model, the low-light image collected in real time is enhanced to obtain an enhanced image.

2. The low-light image enhancement method based on the improved Retinex-Net according to claim 1, characterized in that: In S1, the training set uses the LOL training set and BrightingTrain, and the test set uses the DICM dataset and the LOL evaluation set.

3. The low-light image enhancement method based on the improved Retinex-Net according to claim 1, characterized in that: In S21, the decomposition network is a dual-branch structure, the dual branches include an illumination branch and a reflection branch, the illumination branch uses 7 layers of 7×7 convolutional layers based on the Sigmoid activation function to extract the illumination component (I low ,I normal ); the reflection branch calculates the reflection component (R low ,R normal ), and optimize the reflection component R of the low-light image using 3 layers of 3×3 convolutional layers based on the ReLU activation function low , get the optimized reflection component R low1 .

4. The low-light image enhancement method based on the improved Retinex-Net according to claim 3, characterized in that: In S21, 7 layers of 7×7 convolutional layers based on Sigmoid activation function are used to extract the illumination component (I low ,I normal ) is as follows: I low =Sigmoid(Conv 7×7 (F shallow )); I normal =Sigmoid(Conv 7×7 (F shallow )); Where, F shallow The low-light image S is input using three 3×3 shallow convolutions low With the normal image S normal The calculation formula for the shared basic features obtained after feature extraction is as follows: F shallow =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (S))))); Where, the ReLU function is defined as ReLU(x)=max(0,x); S is S low and S normal .

5. The low-light image enhancement method based on the improved Retinex-Net according to claim 3, characterized in that: In S21, the reflection component (R low ,R normal ) is as follows: R low =S low -I low ; R normal =S normal -I normal ; Optimize the reflection component R of the low-light image using 3 layers of 3×3 convolutional layers based on ReLU activation function low , get the optimized reflection component R low1 The formula is as follows: R low1 =ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (R low ))))))。 6. The low-light image enhancement method based on the improved Retinex-Net according to claim 1, characterized in that: The specific steps of S22 are as follows: S221: Calculate the identity mapping path Identity using the following formula: Identity=Conv 1×1 (R low1 ); S222: Calculate the residual shrinkage path. The specific steps are as follows: First, calculate the residual F(x) using the following formula: F(x)=Conv 3×3 (R low1 ); Secondly, the adaptive threshold T is generated using the following formula: T=σ(W2·ReLU(W1·GAP(F(x)))); Where σ is the Sigmoid activation function; W1 is the weight matrix of the first fully connected layer, with a size of C×C / r, where C is R low1 The number of channels, r is the reduction ratio; W2 is the weight matrix of the second fully connected layer, with a size of C / r×C; GAP represents global average pooling; Then, use the following formula to perform soft threshold denoising to obtain the denoised residual F denoised (x): In the formula, sign takes the symbol ±1; Finally, the reflection component R of the denoised low-light image is calculated using the following formula: denoised : R denoised =ReLU(Identity+F denoised (x))。 7. The low-light image enhancement method based on the improved Retinex-Net according to claim 1, characterized in that: The specific steps of S23 are as follows: S231: Extracting the illumination component I of the low illumination image by downsampling with an encoder low The encoder consists of a four-level downsampling module, each of which contains two 3×3 convolutional layers and a 2×2 maximum pooling layer connected to the convolutional layer. Each convolutional layer uses ReLU activation, and the number of channels in the four levels is gradually doubled from 64 to 512. The encoder formula is as follows: HAVE BEEN l =MaxPool(ReLU(W l,2 *ReLU(W l,1 *HAVE BEEN l-1 ))); Where, E l Represents the output of the encoder level l; MaxPool represents the 2×2 maximum pooling operation for downsampling; W l,1 With W l,2 is the convolution weight corresponding to the two 3×3 convolutions in the lth level; E0=I low , I low is the input illumination component; S232: The decoder symmetrically expands the features extracted by the encoder. The decoder adopts a four-level upsampling structure. Each level first performs a double upsampling through 2×2 transposed convolution to restore the spatial resolution. Then, the channel is spliced ​​with the high-resolution features of the corresponding level of the encoder. Finally, two 3×3 convolutional layers are used to fuse the cross-scale information. Each convolutional layer uses ReLU activation. The number of channels in the four levels decreases from 512 to 64. The decoder formula is as follows: D 4-l =Conv 3×3 (ReLU(Conv 3×3 (Concat(UpSample(D 4-l+1 ),E l )))); Where D 4-l is the output of the (4-l)th level of the encoder, Concat represents the concatenation of the upsampling result and the encoder feature along the channel dimension, and UpSample is the upsampling operation; S233: Calculate the illumination adjustment ΔI using residual learning. The formula is as follows: ΔI=Tanh(W final *D1); Where Tanh is the activation function, W final Represents the weight of the convolutional layer in residual learning; S234: The illumination component I of the low illumination image is converted into low Added to the illumination adjustment value ΔI, the corrected illumination component I is obtained. corrected , the calculation formula is as follows: I corrected =I low +ΔI。 8. The low-light image enhancement method based on the improved Retinex-Net according to claim 1, characterized in that: In S25, the formula of the loss function L is as follows: L=L rec +αL rs +βL is +γL mc ; Among them, L rs =||R1-R h ||1; L mc =||M*exp(-c*M)||1; L rec =||I h -R h *L h ||1+||I1-R1*L1||1; Where: α, β and γ represent the weight parameters of the loss function respectively; R1 is the reflectivity of the low-light image, R h is the reflectivity of the normal image, ||·||1 is the Ll norm; is the illumination gradient of the illumination component of the normal image, is the illumination gradient of the normal image, is the illumination gradient of the illumination component of the low-light image, is the illumination gradient of the low-light image; M is and The sum of , c is the parameter that controls the shape of the function.

9. The low-light image enhancement method based on the improved Retinex-Net according to claim 8, characterized in that: In the loss function, α=0.01, β=0.15, and γ=0.

21.

10. A smart wearable system, characterized by: Under low illumination conditions, the low illumination image enhancement method based on the improved Retinex-Net according to any one of claims 1 to 9 is performed.

Images