Training method and system of low-illumination image enhancement model

Abstract

According to the training method of the low-illumination image enhancement model provided by the invention, the low-illumination image enhancement model is formed through the enhancement network capable of enhancing the brightness of the low-illumination image and the calibration network capable of restoring the brightness of the low-illumination image. And the restored brightness of the calibration network supervises the brightness enhanced by the enhancement network, so that the low-illumination image enhancement model performs self-supervised learning. And inputting the original to-be-enhanced low-illumination image into the low-illumination image enhancement model for training, and obtaining a loss function through the output enhanced image and the original to-be-enhanced low-illumination image. Training parameters of the enhancement network and the calibration network are adjusted in training of the low-illumination image enhancement model through the loss function, and the low-illumination image enhancement model capable of being used for low-illumination image enhancement is obtained. The method adopts an unsupervised learning mode, does not need to depend on a large amount of annotation data, and has higher practicability and popularization value.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and particularly to a training method and system for a low-light image enhancement model. Background Art

[0002] In the field of low-light image enhancement, existing technologies generally face two major challenges: how to ensure computational efficiency while improving image visual quality, and how to maintain the effectiveness of algorithms in unknown and complex real-world scenarios. Traditional methods are mostly based on supervised training. Although these methods improve image quality to a certain extent, they often have problems such as difficulty in parameter adjustment, high computational complexity, and unsuitability for complex scenes.

[0003] In recent years, with the development of deep learning technology, low-light image enhancement methods based on neural networks have gradually become a research hotspot. These methods can automatically learn the rules of image enhancement through training with a large amount of data, and achieve end-to-end image quality improvement. However, existing deep learning-based methods generally have the following problems: high model complexity and large amount of calculation, which makes it difficult to deploy on edge devices; lack of generalization ability for complex scenes, and poor results in practical applications.

[0004] In response to the above problems, this patent proposes a new type of "unsupervised low-light image enhancement automatic learning framework" and a lightweight low-light image enhancement model, aiming to achieve fast, flexible and robust low-light image enhancement. At the same time, a dynamic range fusion exposure suppression algorithm is proposed to deal with the problem of strong light noise that may occur at night. This algorithm is designed to reduce the impact of factors such as night street lights and urban neon lights on image quality, thereby providing clearer visual effects. By cascading the low-light image enhancement model and the dynamic range fusion exposure suppression algorithm, this patent can significantly improve the visual quality and computational efficiency of image enhancement while keeping the model lightweight. In addition, this patent also has good generalization capabilities and can adapt to the needs of different complex scenes.

[0005] Existing low-light image enhancement technologies can be roughly divided into two categories: traditional low-light image enhancement algorithms based on histogram equalization and supervised learning algorithms based on deep learning.

[0006] Traditional low-light image enhancement algorithm based on histogram equalization:

[0007] Histogram equalization (HE) is one of the most commonly used image enhancement methods in the field of image processing. This algorithm redistributes the pixel values ​​of an image according to the distribution, making it approximately uniformly distributed. In low-brightness images, pixels generally appear in a narrow grayscale range, while in images with normal brightness, the histogram tends to be evenly distributed. Therefore, by adjusting the distribution of the histogram of a low-light image, the image contrast can be improved and the image brightness can be restored to a certain extent.

[0008] The histogram equalization based method has the advantages of easy implementation and fast processing speed, but it also has the following problems: (1) chromatic aberration is prone to occur; (2) image details are lost; (3) low brightness areas cannot be enhanced accurately, and normal areas are prone to overexposure.

[0009] Supervised learning algorithms based on deep learning: They use a large amount of paired data for training and can automatically learn the rules of image enhancement. However, such methods rely on a large amount of labeled data and have limited generalization capabilities. Models can often only perform well in scenarios that are consistent with the training data, but are difficult to adapt to unknown scenarios.

[0010] In the field of low-light image enhancement, although existing technologies have made some progress, they still face multiple challenges. First, existing low-light image enhancement methods often sacrifice computational efficiency while improving image visual quality, making these methods difficult to deploy in real-time applications or edge devices. Second, most methods lack sufficient generalization capabilities, especially in unknown and complex real-world scenarios, where their performance usually drops significantly.

[0011] Specifically, existing low-light image enhancement technologies can be divided into two categories: traditional methods based on histogram equalization and supervised learning methods based on deep learning. Although histogram equalization is simple and easy to implement, it is prone to problems such as color aberration, loss of details, and overexposure. Although supervised learning methods based on deep learning can automatically learn the rules of image enhancement, they rely on a large amount of labeled data and have limited generalization capabilities when facing unknown scenes.

[0012] In recent years, with the development of deep learning technology, although some deep learning models have achieved remarkable results in image enhancement, these models are often complex in structure and have a large number of parameters, resulting in large amount of calculation and high memory usage, making them difficult to deploy in practical applications. For example, although the mainstream super-resolution model has excellent performance, due to its increased depth and width, the model parameters are too large and the calculation complexity is too high, making it difficult to apply in mobile devices and embedded systems.

[0013] In addition, existing low-light image enhancement methods lack adaptability to complex scenes. In practical applications, images in low-light environments often contain various noise and interference factors, such as night street lights, city neon lights, etc., which have a negative impact on image quality. However, existing methods often cannot effectively handle these noise and interference factors, resulting in poor quality of enhanced images. Summary of the invention

[0014] The purpose of the present invention is to provide a training method and system for a low-light image enhancement model, aiming to solve the problem in the prior art of sacrificing computational efficiency while improving the visual quality of images, as well as limited generalization ability. By introducing an unsupervised learning framework and lightweight model design, the present invention can significantly improve the visual quality and computational efficiency of image enhancement while keeping the model lightweight, and has good generalization ability to meet the needs of different complex scenes.

[0015] In order to solve the above problems, the present invention proposes a novel "unsupervised low-light image enhancement model automatic learning framework" and a lightweight low-light image enhancement model. The framework uses an unsupervised learning method to automatically learn the rules of image enhancement through the design of a lightweight low-light image enhancement model, without relying on a large amount of labeled data. At the same time, the framework also introduces a dynamic range fusion exposure suppression algorithm, which can effectively deal with the strong light noise problem that may occur at night and improve the visual effect of image enhancement. Through a large number of experimental verifications, the method proposed in the present invention can not only improve the image quality while keeping the model lightweight, but also has good generalization ability and can adapt to the needs of different complex scenes.

[0016] A first aspect of the present invention provides a method for training a low-light image enhancement model, comprising:

[0017] The low-light image enhancement model is composed of an enhancement network that can enhance the brightness of low-light images and a calibration network that can restore the brightness of low-light images. The input of the enhancement network is the original low-light image to be enhanced. The calibration network outputs the enhanced image. The restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model can learn self-supervised.

[0018] The original low-light image to be enhanced is input into the low-light image enhancement model training, and the loss function is obtained by the output enhanced image and the original low-light image to be enhanced.

[0019] In the training of the low-light image enhancement model, the training parameters of the enhancement network and the calibration network are adjusted through the loss function until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

[0020] A second aspect of the present invention provides a training system for a low-light image enhancement model, which includes: an architecture unit of the low-light image enhancement model, a loss function acquisition unit and a training unit, wherein.

[0021] The architecture unit of the low-light image enhancement model is configured to form a low-light image enhancement model through an enhancement network capable of enhancing the brightness of the low-light image and a calibration network capable of restoring the brightness of the low-light image. The input of the enhancement network is the original low-light image to be enhanced. The calibration network outputs the enhanced image. The restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model can learn by self-supervision.

[0022] The loss function acquisition unit is configured to input the original low-light image to be enhanced into the low-light image enhancement model training, and obtain the loss function through the output enhanced image and the original low-light image to be enhanced.

[0023] The training unit is configured to adjust the training parameters of the enhancement network and the calibration network in the training of the low-light image enhancement model through the loss function until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

[0024] A third aspect of the present invention provides a low-light image enhancement model, which is obtained by any low-light image enhancement model training method in the present invention.

[0025] The fourth aspect of the present invention provides a low-light image enhancement method, which enhances the original low-light image to be processed by the low-light image enhancement model in the present invention. The characteristics, technical features, advantages and implementation methods of the training method and system of the low-light image enhancement model will be further explained in a clear and easy-to-understand manner with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart for illustrating a training method of a low-light image enhancement model in one embodiment of the present invention.

[0027] Figure 2 It is a schematic diagram of a learning framework for illustrating a training method of a low-light image enhancement model in one embodiment of the present invention.

[0028] Figure 3 It is a schematic diagram for illustrating the composition of a low-light image enhancement model in one embodiment of the present invention.

[0029] Figure 4 It is a schematic diagram for illustrating the composition of a training system for a low-light image enhancement model in one embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to have a clearer understanding of the technical features, purposes and effects of the invention, the specific embodiments of the present invention are now described with reference to the accompanying drawings. In the drawings, the same reference numerals represent components with the same structure or similar structures but the same functions.

[0031] In this article, "schematic" means "serving as an example, instance or illustration", and any diagram or implementation described as "schematic" in this article should not be interpreted as a more preferred or more advantageous technical solution. In order to make the drawings concise, each figure only schematically shows the parts related to the exemplary embodiment, and they do not represent the actual structure and true proportion of the product.

[0032] like Figure 1 As shown, the first aspect of the present invention provides a training method for a low-light image enhancement model, which includes:

[0033] Step S101, a low-light image enhancement model is formed by an enhancement network capable of enhancing the brightness of a low-light image and a calibration network capable of restoring the brightness of a low-light image. The input of the enhancement network is the original low-light image to be enhanced. The calibration network outputs the enhanced image. The restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model can learn through self-supervision.

[0034] Step S102: input the original low-light image to be enhanced into the low-light image enhancement model training, and obtain the loss function through the output enhanced image and the original low-light image to be enhanced.

[0035] Step S103, in the training of the low-light image enhancement model, the training parameters of the enhancement network and the calibration network are adjusted through the loss function until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

[0036] like Figure 2 As shown in FIG, the present invention proposes an innovative unsupervised low-light image enhancement model automatic learning framework, and its overall architecture is as follows Figure 1 The core idea of ​​this framework is to achieve unsupervised image enhancement learning through two competing networks - Enhance Network and Calibrate Network - and a carefully designed loss function.

[0037] During the training phase, the framework only accepts low-light images as input. First, the enhancement network is responsible for enhancing the input low-light images to brighten them and improve the visual quality of the images. This step is designed to simulate the adaptive adjustment of the human visual system to low-light environments, revealing more details by improving the brightness and contrast of the image.

[0038] Subsequently, the calibration network performs the reverse processing on the enhanced image, that is, darkening it, to obtain the final output image (Output). The purpose of this step is to build a self-supervised learning cycle in which the enhancement network and the calibration network continuously optimize their own performance in the game with each other.

[0039] The input of the loss function is the original low-light image (Input) and the output image (Output) after being processed by the calibration network. By comparing the difference between the two, the loss function can calculate the performance of the current model and adjust the parameters of the enhancement network and the calibration network accordingly. As the training process progresses, the loss function gradually converges, and the overall framework will reach a Nash equilibrium state, at which time the enhancement network and the calibration network achieve an optimal balance in performance.

[0040] During the testing phase, only the trained enhancement network needs to be used. By inputting low-light images into the enhancement network, enhanced images with significantly improved visual quality can be obtained. This unsupervised learning method not only avoids the reliance on a large amount of labeled data, but also effectively improves the generalization ability of the model through the game learning mechanism, enabling it to adapt to various complex low-light environments.

[0041] In one implementation of the training method of the low-light image enhancement model of the present invention, the low-light image enhancement model includes: multiple convolutional layers, and a ReLU activation function is set at the output end of each convolutional layer.

[0042] In another embodiment of the training method of the low-light image enhancement model of the present invention, the convolution layer is three 3×3 convolution layers. The low-light image enhancement model also includes a batch normalization layer and an activation layer. The batch normalization layer is arranged between the multiple convolution layers. The activation layer is arranged at the output end of the low-light image enhancement model.

[0043] like Figure 3 As shown, the present invention uses three 3×3 convolutional layers, two ReLu activation layers, one BN layer and one Sigmoid activation layer. The model structure is extremely lightweight. Compared with the existing models, the model has significant advantages in reasoning speed, model parameter quantity and computational efficiency. The model is combined with the automatic learning framework of the unsupervised low-light image enhancement model to improve the model effect.

[0044] In this paper, a lightweight low-light image enhancement model is proposed to achieve fast, flexible and robust low-light image enhancement. Compared with the traditional deep neural network model, the model significantly reduces the complexity and computational complexity of the model while maintaining high performance, making it more practical and more efficient to deploy in practical applications, especially on real-time or edge devices.

[0045] The lightweight model is mainly composed of three 3×3 convolutional layers, each of which is followed by a ReLU activation function to enhance the nonlinear expression ability of the model. In addition, in order to further improve the performance of the model, a batch normalization (BN) layer is introduced into the model to accelerate model training and enhance the generalization ability of the model. This design enables the model to greatly reduce the number of parameters and computational complexity while maintaining high performance. Specifically, the model is significantly superior to existing models in inference speed, number of model parameters, and computational efficiency, thereby enabling fast and effective low-light image enhancement in various complex scenes. The lightweight low-light image enhancement model of is combined with the automatic learning framework of the unsupervised low-light image enhancement model to form a complete low-light image enhancement system.

[0046] In this system, the unsupervised learning framework is responsible for learning image enhancement rules from a large number of unlabeled low-light images, while the lightweight model is responsible for applying these rules to specific low-light image enhancement tasks. Through unsupervised learning, the system does not need to rely on a large amount of labeled data, thereby improving the practicality and promotion value of the system.

[0047] During the training process, the unsupervised learning framework continuously optimizes the performance of the model through game learning between the enhancement network and the calibration network. The lightweight model learns and accumulates the rules of image enhancement by continuously receiving the enhanced images output by the enhancement network. This training method enables the lightweight model to maintain its lightweight while also having strong generalization capabilities and being able to adapt to various complex low-light environments.

[0048] During the testing phase, only the trained lightweight model needs to be used to enhance low-light images. Due to the simple structure and small amount of computation of the model, it can be efficiently deployed in real-time applications or edge devices to achieve fast and effective low-light image enhancement.

[0049] The lightweight low-light image enhancement model proposed in this paper is not only simple in structure and computationally small, but also achieves high-performance low-light image enhancement by combining it with an unsupervised learning framework. The system does not need to rely on a large amount of labeled data, has high practicality and promotion value, and can achieve fast and effective image enhancement in various complex low-light environments, providing strong support for the development of image processing and computer vision.

[0050] In another embodiment of the training method of the low-light image enhancement model of the present invention, the enhancement network obtains the current brightness enhanced image by fusing the exposure suppression algorithm, the original low-light image to be enhanced and the preliminary brightness enhanced image through dynamic range.

[0051] Dynamic range fusion exposure suppression algorithms include:

[0052] More weight is assigned to the edge information in the original low-light image to be enhanced and the preliminary brightness enhanced image through a local Laplacian filter.

[0053] The saturation of each pixel is obtained according to the standard deviation within the RGB channels of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image.

[0054] The brightness of each pixel is obtained according to the exposure of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image.

[0055] The saturation and brightness of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image, and the corresponding weights are obtained.

[0056] The weights of the pixels at the same position in the original low-light image to be enhanced and the preliminary brightness enhanced image are normalized, and image fusion processing is performed.

[0057] In the present invention, the dynamic range fusion exposure suppression algorithm aims to effectively fuse the original low-light image with the enhanced high-light image to achieve the effect of balanced exposure. The algorithm includes the following steps:

[0058] Step 1: Texture Weight C

[0059] The texture weight C is calculated using a local Laplacian filter, which is a second-order derivative filter for edge detection and can assign more weight to edge information in the image. Its formula can be expressed as Formula 1:

[0060] C(x,y)=|Laplace(I(x,y))| Formula 1

[0061] Wherein, (I(x, y)) is the pixel value of the image at position ((x, y)), and Laplace represents the Laplace filtering operation.

[0062] Step 2: Saturation S

[0063] Saturation (S) is obtained by calculating the standard deviation within the RGB channels of each pixel. Its formula 2 is:

[0064]

[0065] Among them, Ii is the color value of the pixel in the three RGB channels respectively, and μ is the average value of the color values ​​of these three channels.

[0066] Step 3: Brightness E

[0067] The calculation of brightness E takes into account the exposure of the pixel value. For images normalized to the range of 0 to 1, the distribution of brightness weights is based on Gaussian distribution, and its formula 3 is:

[0068]

[0069] Where σ is the standard deviation of the Gaussian distribution and (I(x,y)) is the normalized pixel value.

[0070] Step 4: Weight calculation After obtaining the above three indicators, the weight corresponding to each pixel can be calculated, and its formula 4 is:

[0071] W(x,y)=C(x,y)·S(x,y)·E(,y) Formula 4

[0072] Step 5: Weight normalization and fusion Finally, the weights of pixels at the same position in different images are normalized and fused. Normalized weight W norm The calculation formula 5 of (x, y) is:

[0073]

[0074] Among them, k represents different images. Through the above steps, a balanced exposure image can be obtained, which combines the advantages of the original low-light image and the enhanced high-light image.

[0075] like Figure 4 As shown, in the second aspect of the present invention, a training system for a low-light image enhancement model is provided, which includes: an architecture unit 101 of the low-light image enhancement model, a loss function acquisition unit 201 and a training unit 301, wherein.

[0076] The architecture unit 101 of the low-light image enhancement model is configured to form a low-light image enhancement model through an enhancement network capable of enhancing the brightness of the low-light image and a calibration network capable of restoring the brightness of the low-light image. The input of the enhancement network is the original low-light image to be enhanced. The calibration network outputs the enhanced image. The restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model can be self-supervised for learning.

[0077] The loss function acquisition unit 201 is configured to input the original low-light image to be enhanced into the low-light image enhancement model training, and obtain the loss function through the output enhanced image and the original low-light image to be enhanced.

[0078] The training unit 301 is configured to adjust the training parameters of the enhancement network and the calibration network in the training of the low-light image enhancement model through the loss function until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

[0079] In another embodiment of the training system of the low-light image enhancement model of the present invention, the low-light image enhancement model includes: multiple convolutional layers, and a ReLU activation function is set at the output end of each convolutional layer.

[0080] In another embodiment of the training system of the low-light image enhancement model of the present invention, the convolution layer is three 3×3 convolution layers. The low-light image enhancement model also includes a batch normalization layer and an activation layer. The batch normalization layer is arranged between the multiple convolution layers. The activation layer is arranged at the output end of the low-light image enhancement model.

[0081] In another embodiment of the training system of the low-light image enhancement model of the present invention, the enhancement network obtains the current brightness enhanced image by fusing the exposure suppression algorithm, the original low-light image to be enhanced and the preliminary brightness enhanced image through dynamic range.

[0082] Dynamic range fusion exposure suppression algorithms include:

[0083] More weight is assigned to the edge information in the original low-light image to be enhanced and the preliminary brightness enhanced image through a local Laplacian filter.

[0084] The saturation of each pixel is obtained according to the standard deviation within the RGB channels of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image.

[0085] The brightness of each pixel is obtained according to the exposure of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image.

[0086] The saturation and brightness of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image, and the corresponding weights are obtained.

[0087] The weights of the pixels at the same position in the original low-light image to be enhanced and the preliminary brightness enhanced image are normalized, and image fusion processing is performed.

[0088] A third aspect of the present invention provides a low-light image enhancement model, which is obtained by any low-light image enhancement model training method in the present invention.

[0089] A fourth aspect of the present invention provides a low-light image enhancement method, which enhances the original low-light image to be processed by using the low-light image enhancement model of the present invention.

[0090] The "unsupervised low-light image enhancement model automatic learning framework", lightweight low-light image enhancement model and dynamic range fusion exposure suppression algorithm proposed in this invention can not only improve image quality while keeping the model lightweight, but also have good generalization ability and can adapt to the needs of different complex scenes. In addition, the technology of this invention adopts an unsupervised learning method and does not need to rely on a large amount of labeled data, which has stronger practicality and promotion value.

[0091] On the one hand, the present invention uses an unsupervised low-light image enhancement model automatic learning framework and an unsupervised learning method, without relying on a large amount of labeled data, and the model can automatically learn the rules of image enhancement. This can not only improve the practicality of the model and reduce the dependence on data annotation, but also achieve high-quality image enhancement while keeping the model lightweight.

[0092] On the other hand, the present invention adopts a lightweight design and significantly reduces the computational complexity and memory usage of the model by optimizing the model structure and parameters. This enables the model to be efficiently deployed in real-time applications or edge devices to meet the needs of fast image enhancement. On the other hand, the present invention introduces a dynamic range fusion exposure suppression algorithm to address the problem of strong light noise that may occur at night. The algorithm can effectively handle strong light interference at night, reduce its impact on image quality, and thus improve the visual effect of the enhanced image.

[0093] The low-light image enhancement model in the present invention has the following overall technical effects through testing:

[0094] Test equipment: NVIDIA GeForce RTX 3070Laptop GPU

[0095] Test image size: 1920 x 1080

[0096] (1) Frame rate index: 218.89fps

[0097] (2) Storage space occupied by the model: 5.13kb

[0098] (3) Model calculation amount: 0.5038848 GFlops

[0099] (4) Number of model parameters: 252.

[0100] It should be understood that although this specification is described according to various implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0101] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A training method for a low-light image enhancement model, characterized in that: It includes: The low-light image enhancement model is formed by an enhancement network capable of enhancing the brightness of the low-light image and a calibration network capable of restoring the brightness of the low-light image; The input of the enhancement network is the original low-light image to be enhanced; the calibration network outputs the enhanced image; the restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model can learn through self-supervision; Inputting the original low-light image to be enhanced into the low-light image enhancement model training, and obtaining a loss function through the output enhanced image and the original low-light image to be enhanced; The loss function is used to adjust the training parameters of the enhancement network and the calibration network during the training of the low-light image enhancement model until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

2. The training method of the low-light image enhancement model according to claim 1, characterized in that: The low-light image enhancement model includes: multiple convolutional layers, and a ReLU activation function is set at the output end of each convolutional layer.

3. The training method of the low-light image enhancement model according to claim 2, characterized in that: The convolution layer is three 3×3 convolution layers; the low-light image enhancement model also includes a batch normalization layer and an activation layer; the batch normalization layer is arranged between the multiple convolution layers; the activation layer is arranged at the output end of the low-light image enhancement model.

4. The training method for the low-light image enhancement model according to claim 1, characterized in that: The enhancement network obtains a current brightness enhanced image by fusing the exposure suppression algorithm, the original low-light image to be enhanced and the preliminary brightness enhanced image through dynamic range; The dynamic range fusion exposure suppression algorithm includes: Allocating more weights to edge information in the original low-light image to be enhanced and the preliminary brightness enhanced image by using a local Laplacian filter; Obtaining the saturation of each pixel according to the standard deviation in the RGB channel of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image; Obtaining the brightness of each pixel according to the exposure of each pixel in the original low-light image to be enhanced and the preliminary brightness-enhanced image; Obtaining the saturation and brightness of each pixel in the original low-light image to be enhanced and the preliminary brightness-enhanced image, and the corresponding weights; The weights of the pixels at the same position in the original low-light image to be enhanced and the preliminary brightness enhanced image are normalized, and image fusion processing is performed.

5. A training system for a low-light image enhancement model, characterized in that: It includes: The architecture unit, loss function acquisition unit and training unit of the low-light image enhancement model, wherein; The architecture unit of the low-light image enhancement model is configured to form the low-light image enhancement model through an enhancement network capable of enhancing the brightness of the low-light image and a calibration network capable of restoring the brightness of the low-light image; the input of the enhancement network is the original low-light image to be enhanced; the calibration network outputs the enhanced image; the restored brightness of the calibration network supervises the enhanced brightness of the enhancement network, so that the low-light image enhancement model is self-supervised learning; A loss function acquisition unit, configured to input the original low-light image to be enhanced into the low-light image enhancement model training, and obtain a loss function through the output enhanced image and the original low-light image to be enhanced; The training unit is configured to adjust the training parameters of the enhancement network and the calibration network in the training of the low-light image enhancement model through the loss function until the low-light image enhancement model converges, thereby obtaining a low-light image enhancement model that can be used for low-light image enhancement.

6. The training system for the low-light image enhancement model according to claim 5, characterized in that: The low-light image enhancement model includes: multiple convolutional layers, and a ReLU activation function is set at the output end of each convolutional layer.

7. The training system for the low-light image enhancement model according to claim 6, characterized in that: The convolution layer is three 3×3 convolution layers; the low-light image enhancement model also includes a batch normalization layer and an activation layer; the batch normalization layer is arranged between the multiple convolution layers; the activation layer is arranged at the output end of the low-light image enhancement model.

8. The training system for the low-light image enhancement model according to claim 5, characterized in that: The enhancement network obtains a current brightness enhanced image by fusing the exposure suppression algorithm, the original low-light image to be enhanced and the preliminary brightness enhanced image through dynamic range; The dynamic range fusion exposure suppression algorithm includes: Allocating more weights to edge information in the original low-light image to be enhanced and the preliminary brightness enhanced image by using a local Laplacian filter; Obtaining the saturation of each pixel according to the standard deviation in the RGB channel of each pixel in the original low-light image to be enhanced and the preliminary brightness enhanced image; Obtaining the brightness of each pixel according to the exposure of each pixel in the original low-light image to be enhanced and the preliminary brightness-enhanced image; Obtaining the saturation and brightness of each pixel in the original low-light image to be enhanced and the preliminary brightness-enhanced image, and the corresponding weights; The weights of the pixels at the same position in the original low-light image to be enhanced and the preliminary brightness enhanced image are normalized, and image fusion processing is performed.

9. A low-light image enhancement model, characterized in that: The low-light image enhancement model is obtained by the training method of any one of claims 1 to 4 for a low-light image enhancement model.

10. A low-light image enhancement method, characterized in that: The original low-light image to be processed is enhanced by the low-light image enhancement model in claim 9.

Images