A Low-Light Image Enhancement Method Based on Improved Retinex-Net and a Smart Wearable System

Abstract

The application provides a low-illumination image enhancement method and an intelligent wearable system based on an improved Retinex-Net, wherein the image enhancement 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; designing an improved Retinex-Net-based low-illumination image enhancement model and training and verifying the model to obtain a trained low-illumination image enhancement model, wherein the low-illumination image enhancement model is composed of a decomposition network, an adjustment network and a reconstruction network, and a subnetwork formed by an identity mapping and a residual shrinkage construction unit is added to a reflection path of the decomposition network; and the trained low-illumination image enhancement model is used to enhance an image to obtain an enhanced image. The low-illumination image enhancement method and the intelligent wearable system can effectively suppress noise generated in the image decomposition process, improve the definition and quality of the image, and realize low-illumination image enhancement.

Description

Technical Field

[0001] This invention relates to the field of image enhancement and smart wearable system technology, and in particular to a low-light image enhancement method and smart wearable system based on an improved Retinex-Net. Background Technology

[0002] With advancements in modern transportation technology and stringent safety management systems, accident rates have significantly decreased. Despite overall improved safety, drivers and pilots still face specific risk factors, among which weather conditions pose 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 driving safety. Image enhancement, a key technology for improving image readability, can effectively improve operator visibility through low-light enhancement applications. Traditionally, pilots rely on Instrument Flight Rules (IFR), and drivers must reduce vehicle speed and activate fog lights. However, in extremely low visibility conditions, there are still safety hazards such as inaccurate judgments and delayed reactions. Therefore, intelligent visibility enhancement technology has become an industry necessity. The rapid development of artificial intelligence provides new opportunities for this, with its high-precision processing and all-weather operation capabilities offering an effective solution to operational challenges in low-visibility environments. This technology has already achieved significant results in intelligent transportation, autonomous driving, and security monitoring, making the development of intelligent vision enhancement systems suitable for multiple scenarios particularly urgent.

[0003] However, while artificial intelligence is currently being applied in the transportation sector, dedicated solutions for drivers and pilots remain insufficient. Traditional methods for dealing with low visibility suffer from limitations such as low efficiency and poor adaptability, and existing augmentation technologies struggle to meet the real-time requirements in complex environments.

[0004] Therefore, proposing a new method for low-light image enhancement and a smart wearable system to assist pilots and drivers in safe driving 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 a smart wearable system based on an improved Retinex-Net to solve the problems existing in traditional low-light image enhancement methods.

[0006] This 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: Design a low-light image enhancement model based on the improved Retinex-Net and train and validate the low-light image enhancement model using the training set and test set 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. The decomposition network incorporates a sub-network formed by identity mapping and residual shrinkage building units into its reflection path. The training process of the model is as follows:

[0010] S21: Use the decomposition network to process the input low-light image S low Compared with normal image S normal The corresponding decomposition is into reflection components (R) low ,R normal ) and illumination component (I low ,I normal ), and the reflectance component R of the low-light image. low Optimization is performed to obtain the reflectance component R of the optimized low-light image. low1 ;

[0011] S22: Utilizing a subnetwork formed by identity mapping and residual shrinkage building units to optimize the reflectance component R of the low-light image. low1 Noise suppression is performed to obtain the reflectance component R of the denoised low-light image. denoised ;

[0012] S23: Utilize the adjustment network to adjust the illumination component I of the low-light image. low The correction is performed to obtain the corrected illumination component I. corrected ;

[0013] S24: Using the reconstruction network, the denoised reflection component R... denoised With the corrected illumination component I corrected Element-wise multiplication is performed to recover the enhanced image corresponding to the low-light image;

[0014] S25: Calculate the error between the enhanced image and the normal image using the loss function and perform backpropagation to update the parameters of the low-light image enhancement model until the training of the low-light image enhancement model is completed.

[0015] S3: The trained low-light image enhancement model is used to enhance the low-light image acquired in real time to obtain the 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 has a dual-branch structure, the dual branches including an illumination branch and a reflection branch, the illumination branch using 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) through the residual. low ,R normal The reflectance component R of the low-light image is optimized using three 3×3 convolutional layers based on the ReLU activation function. low The optimized reflection component R is obtained. low1 .

[0018] Further optimization, in S21, a 7×7 convolutional layer based on the Sigmoid activation function is used to extract the illumination component (I). low ,I normal The formula is as follows:

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

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

[0021] In the formula, F shallow To process the input low-light image S using three 3×3 shallow convolutions... low Compared with normal image S normal The shared basic features obtained after feature extraction are calculated using the following formula:

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

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

[0024] Further optimization, in S21, the reflection component (R) is calculated through the residual. low ,R normal The formula is as follows:

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

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

[0027] The reflectance component R of the low-light image is optimized using three 3×3 convolutional layers based on the ReLU activation function. low The optimized reflection component R is obtained. 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 optimization, 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] In the formula, 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] Next, soft thresholding denoising is performed using the following formula to obtain the denoised residual F. denoised (x):

[0039]

[0040] In the formula, sign takes the sign ±1;

[0041] Finally, the reflectance 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 optimization, the specific steps of S23 are as follows:

[0044] S231: Extract the illumination component I of the low-light image using encoder downsampling. low The encoder is characterized by comprising a four-level downsampling module, each level containing two 3×3 convolutional layers and a 2×2 max-pooling layer connected to the convolutional layers. Each convolutional layer uses ReLU activation, and the number of channels increases progressively from 64 to 512 across the four levels. The encoder formula is as follows:

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

[0046] In the formula, E l This represents the output of the l-th stage of the encoder; MaxPool represents a 2×2 max pooling operation used for downsampling; W l,1 With W l,2 These are the convolution weights corresponding to the two 3×3 convolutions in level l; E0 = I low I low The input illumination component;

[0047] S232: The decoder symmetrically unfolds the features extracted by the encoder. The decoder employs a four-level upsampling structure. Each level first performs a 2×2 transposed convolution to double the upsampling to restore the spatial resolution, then concatenates the channels with the high-resolution features of the corresponding level of the encoder, and finally fuses cross-scale information through two 3×3 convolutional layers. 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] In the formula, D 4-lThis is the output of the encoder at level (4-1). Concat means concatenating the upsampled result with the encoder feature along the channel dimension, and UpSample is the upsampling operation.

[0050] S233: Calculate the illumination adjustment ΔI using residual learning, as shown in the following formula:

[0051] ΔI=Tanh(W final *D1);

[0052] In the formula, Tanh is the activation function. W final This represents the weights of the convolutional layer in residual learning;

[0053] S234: The illumination component I of the low-light image... low Adding this to the illumination adjustment amount ΔI yields the corrected illumination component I. corrected The calculation formula is as follows:

[0054] I corrected =I low +ΔI.

[0055] Further preferably, in S25, the formula for 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] In the formula: α, β, and γ represent the weight parameters of the loss function; R1 is the reflectance of the low-light image, R h Let ||·||1 be the reflectance of a normal image, and ||·||1 be the L1 norm. The illumination gradient of the illumination component in a normal image. The illumination gradient of a normal image. The illumination gradient of the illumination component in a low-light image. M is the illumination gradient of the low-light image; and The sum of these, where c is a parameter controlling the shape of the function.

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

[0063] The present invention also provides a smart wearable system for performing the above-described low-light image enhancement method based on the improved Retinex-Net in low-light conditions.

[0064] The present invention provides a low-light image enhancement method and intelligent wearable system based on an improved Retinex-Net. By utilizing a low-light image enhancement model based on Retinex-Net, the low-light image and normal image are effectively decomposed into two parts, illumination component and reflection component, through a sub-network formed by identity mapping and residual shrinkage building units. This can effectively suppress noise generated during the image decomposition process, improve the image clarity and quality, and thus achieve low-light image enhancement. Attached Figure Description

[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 Implementation

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

[0067] While artificial intelligence has been applied in the transportation sector, solutions specifically for drivers and pilots remain insufficient. Traditional low-visibility coping methods are inefficient and poorly adaptable, and existing technologies struggle to meet real-time requirements.

[0068] In response to the above problems, such as Figure 1 As shown, this 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, while the test set uses the DICM dataset and the LOL evaluation set.

[0071] To make the results more accurate, a large amount of training data is needed during the learning phase. To address the problem of the limited number of images in the dataset, data augmentation steps are also included. Specifically, data augmentation is performed by randomly flipping, shrinking, and translating the images in the dataset.

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

[0073] like Figure 1 As shown, the low-light image enhancement model consists of three parts: a decomposition network, an adjustment network, and a reconstruction network. The reflection path of the decomposition network incorporates a sub-network formed by identity mapping and residual shrinkage building units. The training process of the model is as follows:

[0074] S21: Use the decomposition network to process the input low-light image S low Compared with normal image S normal The corresponding decomposition is into reflection components (R) low ,R normal ) and illumination component (I low ,I normal ), and the reflectance component R of the low-light image. low Optimization is performed to obtain the reflectance component R of the optimized low-light image. low1 ;

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

[0076] Specifically, a 7×7 convolutional layer with a sigmoid activation function is used to extract the illumination component (I). low ,I normal The formula is as follows:

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

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

[0079] In the formula, F shallow To process the input low-light image S using three 3×3 shallow convolutions... low Compared with normal image S normal The shared basic features obtained after feature extraction are calculated using the following formula:

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

[0081] In the formula, the ReLU function is defined as ReLU(x) = max(0,x), 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 Each pixel value is constrained to [0,1].

[0083] Among them, the reflection component (R) is calculated through the residual. low ,R normal The formula is as follows:

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

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

[0086] The reflectance component R of the low-light image is optimized using three 3×3 convolutional layers based on the ReLU activation function. low The reflectance component R of the optimized low-light image is obtained. low1 The formula is as follows:

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

[0088] In the formula, R low1The reflectance component of the optimized low-light image;

[0089] The reflection component of the normal image is no longer further processed to prevent overprocessing of the normal image from introducing artifacts or loss of detail.

[0090] S22: Utilizing a subnetwork formed by identity mapping and residual shrinkage building units to optimize the reflectance component R of the low-light image. low1 Noise suppression is performed to obtain the reflectance 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] In the formula, 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] Next, soft thresholding denoising is performed using the following formula to obtain the denoised residual F. denoised (x):

[0101]

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

[0103] Finally, the reflectance 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: Utilize the adjustment network to adjust the illumination component I of the low-light image. low The correction is performed to obtain the corrected illumination component I. corrected ;

[0106] S231: Extract the illumination component I of the low-light image using encoder downsampling. low The encoder is characterized by comprising a four-level downsampling module, each level containing two 3×3 convolutional layers and a 2×2 max-pooling layer connected to the convolutional layers. Each convolutional layer uses ReLU activation, and the number of channels increases progressively from 64 to 512 across the four levels. The encoder formula is as follows:

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

[0108] In the formula, E l This represents the output of the encoder's l-th stage; MaxPool represents a 2×2 max pooling operation, used for downsampling to reduce data dimensionality; W l,1 With W l,2 These are the convolution weights corresponding to the two 3×3 convolutions in level l, used to perform convolution operations on the input to extract features; E0 = I low I low The input illumination component;

[0109] S232: The features extracted by the encoder are symmetrically unfolded using the decoder. The decoder adopts a four-level upsampling structure. Each level first performs upsampling by a factor of 2 through a 2×2 transposed convolution to restore the spatial resolution, and then concatenates the channels with the high-resolution features of the corresponding level of the encoder. Finally, it fuses cross-scale information through two 3×3 convolutional layers. The number of channels in the four levels decreases 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] In the formula, D 4-l For the output of the encoder at level (4-1), Concat is the concatenation of the upsampled result and encoder features along the channel dimension, and UpSample is the upsampling operation;

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

[0113] ΔI=Tanh(W final *D1);

[0114] In the formula, Tanh is the activation function. Used to constrain the magnitude of illumination adjustment ΔI, preventing over-adjustment, W final This represents the weights of the convolutional layer in residual learning.

[0115] S234: The illumination component I of the low-light image... low Adding this to the illumination adjustment amount ΔI yields the corrected illumination component I. corrected The calculation formula is as follows:

[0116] I corrected =I low +ΔI;

[0117] S24: Using the reconstruction network, the denoised reflection component R... denoised With the corrected illumination component I corrected Element-wise multiplication recovers 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 backpropagation to update the parameters of the low-light image enhancement model until the training of the low-light image enhancement model is completed.

[0120] The formula for 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] In the formula: α, β, and γ represent the weight parameters of the loss function; R1 is the reflectance of the low-light image, R h Let ||·||1 be the reflectance of a normal image, and ||·||1 be the L1 norm. The illumination gradient of the illumination component in a normal image. The illumination gradient of a normal image. The illumination gradient of the illumination component in a low-light image. M is the illumination gradient of the low-light image; and The sum, where c is a parameter controlling the shape of the function;

[0127] L rec The loss function reflects the accuracy of the decomposition. Since the light and reflection components that are decomposed need to follow the Retinex theory, they can be restored to the original image.

[0128] L rs To account for the reflectance similarity loss, according to previous theories, the reflectance of the image decomposed under different illuminations should be the same;

[0129] L is To account for the loss of illumination smoothness, the gradients in the low-light and normal directions of the illumination component are calculated.

[0130] L mc As a loss of consistency, a large or small illumination gradient between low and normal illumination indicates that the illumination is at the surface or edge of an object in equilibrium. Penalties are only imposed when there is a small difference between the gradients.

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

[0132] S3: Use a trained low-light image enhancement model to enhance the low-light images acquired in real time, and obtain the enhanced image.

[0133] This invention provides a low-light image enhancement method based on an improved Retinex-Net. Addressing the shortcomings of current low-light image enhancement methods in terms of noise generation, this method utilizes a Retinex-Net-based low-light image enhancement model. Through a sub-network formed by identity mapping and residual shrinkage building units, it effectively decomposes low-light and normal images into illumination and reflection components. This effectively suppresses noise generated during image decomposition, improving image clarity and quality, thereby achieving low-light image enhancement. Specifically, Retinex-Net exhibits significant noise in low-light image enhancement, leading to loss of image details. Therefore, introducing a sub-network formed by identity mapping and residual shrinkage building units into the decomposition network removes noise generated during decomposition, improving the model's ability to enhance low-light images.

[0134] The present invention also provides a smart wearable system for performing the above-described low-light image enhancement method based on the improved Retinex-Net in low-light conditions.

[0135] This invention applies the aforementioned low-light image enhancement method based on the improved Retinex-Net to a smart wearable system, providing assistance to drivers or pilots. It offers several advantages: First, the improved identity mapping and residual shrinkage building blocks effectively suppress noise generated during image enhancement, preventing noise from affecting image quality and ensuring a more natural and accurate visual effect after enhancement. Second, the improved Retinex-Net effectively enhances details in low-light images, providing clearer visual information and helping drivers and pilots see their surroundings more clearly at night or in dim environments, improving driving or flight safety. Furthermore, this method effectively enhances images in various low-light environments, providing stable and reliable image enhancement effects in urban nightscapes, rural roads, and inclement weather, enhancing the visual perception capabilities of drivers or pilots.

[0136] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take 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.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0137] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0141] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An improved Retinex-Net-based low-light image enhancement method, characterized in that, The method comprises the following steps: S1: collecting a dataset with a low-illumination label and dividing it into a training set and a test set, wherein the dataset comprises images of different scenes and different illumination conditions; S2: designing an improved Retinex-Net-based low-illumination image enhancement model and training and verifying the low-illumination image enhancement model by using the training set and the test set to obtain a trained low-illumination image enhancement model; The low-illumination image enhancement model is composed of a decomposition network, an adjustment network and a reconstruction network, wherein a subnetwork formed by an identity mapping and a residual shrinkage construction unit is added to the reflection path of the decomposition network, and the training process of the model is as follows: S21: using the decomposition network to decompose the input low-illumination image S low into a reflection component (R low ) and an illumination component (I low ), and optimizing the reflection component R low of the low-illumination image to obtain an optimized reflection component R low1 of the low-illumination image; normal normal normal low low1 ​ S22: using a sub-network formed by an identity mapping and a residual contraction building unit on the optimized low-illumination image reflection component R low1 noise suppression is performed to obtain the denoised low-illumination image reflection component R denoised ; S23: adjusting the illumination component I of the low-illumination image by using the adjustment network low correcting to obtain a corrected illumination component I corrected ; S24: reconstructing the reflection component R denoised with the corrected illumination component I corrected element-wise multiplication, to recover an enhanced image corresponding to the low-light image; S25: calculating the error between the enhanced image and the normal image by 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: enhancing a real-time collected low-illumination image by using the trained low-illumination image enhancement model 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 adopts LOL training set and BrightingTrain, and the test set selects DICM dataset and 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 double-branch structure, the double-branch includes an illumination branch and a reflection branch, the illumination branch adopts a 7-layer 7x7 convolution layer based on a Sigmoid activation function to extract an illumination component (I low , normal );the reflection branch calculates a reflection component (R low , normal ) by residual calculation, and optimizes the reflection component R low of the low-illumination image by using a 3-layer 3x3 convolution layer based on a ReLU activation function to obtain an 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, a 7-layer 7x7 convolutional layer with a sigmoid activation function is used to extract the illumination component (I low ,I normal ) as follows: I low = Sigmoid(Conv 7×7 (F shallow )); I normal = Sigmoid(Conv 7×7 (F shallow )); In the formula, F shallow is to use three 3x3 shallow convolutional layers on the input low-light image S low and normal image S normal After feature extraction, the shared basic features are obtained, and the calculation formula is as follows: F shallow = ReLU(Conv 3×3 (ReLU(Conv 3×3 (ReLU(Conv 3×3 (S)))))) where 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 formula for calculating the reflection component (R low ,R normal ) from the residual is as follows: R low = S low - I low ; R normal = S normal - I normal ; The low-light image is optimized using 3 layers of 3x3 convolution layers with ReLU activation function to optimize the reflection component R of the low-light image low The formula for the optimized reflection component R low1 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: calculating the identity mapping path Identity by using the following formula: Identity = Conv 1×1 (R low1 ); S222: calculating the residual shrinkage path, and the specific steps are as follows: First, calculating the residual F(x) by using the following formula: F(x) = Conv 3×3 (R low1 ); Second, generating an adaptive threshold T by using the following formula: T=σ(W2·ReLU(W1·GAP(F(x)))); where σ is a Sigmoid activation function; W1 is a weight matrix of the first fully connected layer, with a size of C x C / r, where C is the number of channels of R low1 and r is a reduction ratio; W2 is a weight matrix of the second fully connected layer, with a size of C / r x C; GAP denotes a global average pooling; Then, the soft threshold denoising is performed by using the following formula to obtain the denoised residual F denoised (x): In the formula, sign takes the sign ±1. Finally, the reflectance component R of the de-noised low-illumination image is calculated using the following equation 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 an illumination component I of the low-illumination image using an encoder downsampling low characterized in that the encoder is composed of four downsampling modules, each of which contains two 3x3 convolution layers and a 2x2 max-pooling layer connected to the convolution layers, wherein each convolution layer uses ReLU activation, and the number of channels increases from 64 to 512 step by step, and the encoder formula is as follows: E l = MaxPool(ReLU(W l,2 * ReLU(W l,1 * E l-1 ))) ; In the formula, E l This represents the output of the l-th stage of the encoder; MaxPool represents a 2×2 max pooling operation used for downsampling; W l,1 With W l,2 These are the convolution weights corresponding to the two 3×3 convolutions in level l; E0 = I low I low The input illumination component; S232: using a decoder to extract the features extracted by the encoder, wherein the decoder adopts a four-level upsampling structure, each level first performs two times upsampling by using 2x2 transpose convolution to recover the spatial resolution, then performs channel splicing with the high-resolution features of the corresponding level of the encoder, and finally fuses the cross-scale information through two 3x3 convolution layers, each convolution layer uses ReLU activation, the channel number of the four levels decreases from 512 to 64, and the formula of the decoder is as follows: D 4-l = Conv 3×3 (ReLU(Conv 3×3 (Concat(UpSample(D 4-l+1 ), E l )))) In the formula, D 4-l is the output of the (4-l)-th stage of the encoder, Concat represents concatenating the up-sampling result and the encoder feature along the channel dimension, and UpSample is an up-sampling operation. S233: calculating the illumination adjustment amount ΔI by using residual learning, and the formula is as follows: ΔI = Tanh(W final * D1); In the formula, Tanh is an activation function, W final denotes the weights of the convolutional layer in the residual learning; S234: add the illumination adjustment amount ΔI to the illumination component I of the low-illumination image to obtain a corrected illumination component I low S234: add the illumination adjustment amount ΔI to the illumination component I of the low-illumination image to obtain a corrected illumination component I corrected , and the calculation formula is as follows: I corrected = I low + ΔI. 8.The low-light image enhancement method based on the improved Retinex-Net of 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 ; wherein 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; In the formula: α, β, and γ represent the weight parameters of the loss function; R1 is the reflectance of the low-light image, R h Let ||·||1 be the reflectance of a normal image, and ||·||1 be the L1 norm. The illumination gradient of the illumination component in a normal image. The illumination gradient of a normal image. The illumination gradient of the illumination component in a low-light image. M is the illumination gradient of the low-light image; and The sum of these, where c is a parameter controlling 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. An intelligent wear system, characterized in that: Performing the improved Retinex-Net-based low-illumination image enhancement method of any one of claims 1 to 9 under a low-illumination condition.

Images