Semi-supervision-based low-illumination image enhancement method and related equipment

Abstract

The invention discloses a semi-supervision-based low-illumination image enhancement method and related equipment, and relates to the field of digital image processing. The low-illumination image enhancement method based on semi-supervision comprises the following steps: constructing a dimming module based on a mixed intensity network, and simulating color distortion and darkening processes of an image under a low-illumination condition based on the dimming module; converting the high-quality image into a composite image under a low-light condition based on a dimming module; training an image enhancement model of the UNet architecture in an unsupervised mode based on the manual training data set; the image enhancement model is trained based on an artificial training data set so as to improve the brightness recovery level of a low-illumination image; realizing image enhancement of the low-illumination image based on an image enhancement model; a brightness value can be set in the image enhancement process; a brightness level of the output image is controlled based on the brightness value. And the brightness value is introduced as condition input, so that image distortion caused by excessive enhancement is avoided, and the naturalness and applicability of an enhancement result are improved.

Description

Technical Field

[0001] The present application relates to the field of digital image processing, and in particular to a semi-supervised low-light image enhancement method and related equipment. Background Art

[0002] Low-light image enhancement is a challenging task as it requires increasing image brightness while retaining natural colors. The main challenges in this process include the differences in how cameras process images and the difficulty in obtaining paired image data under realistic lighting conditions.

[0003] Existing low-light image enhancement methods are either fully supervised or unsupervised; for fully supervised methods, although they can achieve good results on paired datasets, they perform poorly when processing unseen data (such as images from different cameras). Meanwhile, unsupervised methods, although they have better generalization capabilities, often only provide lower quality image enhancement results. Summary of the invention

[0004] To solve the above problems, the present application discloses a semi-supervised low-light image enhancement method, which generates high-quality synthetic low-light images through a dimming module; introduces dimming factors and brightness values ​​as conditional inputs to avoid image distortion caused by excessive enhancement, and improves the naturalness and applicability of the enhancement results.

[0005] The first technical solution adopted in this application is: providing a low-light image enhancement method based on semi-supervision, comprising the following steps: Constructing a dimming module based on a hybrid intensity network, and simulating the color distortion and dimming process of an image under low light conditions based on the dimming module; Converting the high-quality image into a synthetic image under low-light conditions based on the dimming module to generate an artificial training dataset; The image enhancement model of the UNet architecture is trained in an unsupervised manner based on the artificial training data set; the image enhancement model is trained based on the artificial training data set to improve the brightness restoration level of the low-light image; Image enhancement of low-light images is achieved based on the image enhancement model; a brightness value can be set during the image enhancement process; and the brightness level of the output image is controlled based on the brightness value.

[0006] Furthermore, constructing the dimming module includes the following steps: The i-th channel of the reflectivity is calculated by channel-by-channel normalization. The formula is: ; in, represents Hadamard (element-wise) partitioning, and denote the i-th channel of image I and reflectivity respectively; Adjust the reflection component R based on the mixture density network (MDN), which is based on the input reflection component R, the illumination component L, and the dimmed illumination component , output the adjusted reflection component ; Based on the Hadamard product, the adjusted reflection component and the dimmed light component Combine to get the final dark image .

[0007] Furthermore, the mixed density network adjusting the reflection component R comprises the following steps: Modeling reflectivity distortion in low-light images; Independently process a 5-dimensional representation of each pixel, which includes reflectance, illumination component L, and dimmed illumination component ; The reflectivity includes three color channels; Outputs the mean, standard deviation, and mixing coefficient of each color channel, describing the probability distribution of color distortion.

[0008] Furthermore, the generation of the artificial training data set includes the following steps: Use high-quality images from public datasets as input; Call the dimming module to convert the high-quality image into a synthetic image under low-light conditions; The high-quality images are paired with the corresponding low-light synthetic images to form training data pairs; multiple training data pairs constitute an artificial training dataset.

[0009] Furthermore, the dimming module can set a dimming factor when converting a high-quality image into a synthetic image under low-light conditions; the dimming factor is used to control the illumination level for generating a low-light image, and different dimming factor values ​​are set to generate low-light images with different illumination levels.

[0010] Furthermore, it also includes fine-tuning the image enhancement model of the UNet architecture based on real images to improve the adaptability of the UNet architecture to different cameras.

[0011] Furthermore, the UNet architecture includes a brightness setting module, and the setting of the brightness value is implemented based on the brightness setting module.

[0012] Furthermore, the UNet architecture receives a dimmed image, a dimmed image after histogram equalization, a color map, and lighting information as input, and outputs a residual map in combination with the brightness value set by the brightness setting module, and the residual map is combined with the dimmed image to obtain a final enhanced image.

[0013] The second technical solution adopted in the present application is: an electronic device is provided, comprising: a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory to implement the semi-supervised low-light image enhancement method as described in any of the above items.

[0014] The third technical solution adopted in the present application is: providing a computer-readable storage medium, wherein the computer-readable storage medium stores program data, and the program data can be executed by a processor to implement the semi-supervised low-light image enhancement method as described in any of the above items.

[0015] Due to the adoption of the above technical solution, the present application has at least one of the following beneficial effects compared with the prior art:

[0016] 1. Introducing brightness value as conditional input avoids image distortion caused by over-enhancement and improves the naturalness and applicability of the enhancement results.

[0017] 2. The UNet architecture is initially trained in an unsupervised manner and fine-tuned using a small number of real image pairs so that the model can better adapt to the imaging characteristics of different cameras.

[0018] 3. By constructing a dimming module, a large number of paired samples of synthetic low-light images and high-quality images are generated, which effectively solves the problem of scarcity of real paired data.

[0019] 4. Combining the mixed density network to model the color distortion and reflectivity changes of low-light images, and the residual learning mechanism of the UNet architecture, high-quality low-light image enhancement is achieved, which effectively preserves the natural color and detail information of the image and significantly improves the image quality under low-light conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 A schematic diagram of a process flow of an embodiment of a semi-supervised low-light image enhancement method provided in the present application; Figure 2 A schematic diagram of the process of training a UNet model for a dimming module provided in the present application; Figure 3 A schematic diagram of a process of creating a dark image by a dimming module provided in the present application; Figure 4This is a schematic diagram of the structure of an embodiment of a computer device of the present application; Figure 5 This is a schematic diagram of the structure of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It will be appreciated that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some but not all structures related to the present application are shown in the drawings. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the art without making creative work are within the scope of protection of the present application.

[0022] The terms "first", "second", etc. in this application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0023] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0024] Existing low-light image enhancement methods include fully supervised or unsupervised methods. Fully supervised low-light image enhancement methods have the disadvantages of high dependence on paired data, insufficient model generalization ability, and difficulty in obtaining paired data. Unsupervised low-light image enhancement methods have the disadvantage of insufficient enhancement effect quality while ensuring generalization ability. In order to solve the above problems, the present application provides a low-light image enhancement method based on semi-supervision, such as Figure 1 As shown, Figure 1 The flowchart of an embodiment of a semi-supervised low-light image enhancement method provided in this application includes the following steps:

[0025] Step S11: construct a dimming module based on a hybrid intensity network, and simulate the color distortion and dimming process of an image under low-light conditions based on the dimming module; existing pairs of high-quality images and their corresponding low-light images are not easy to obtain, and the lighting effect of low-light images is single, and it is impossible to obtain multiple groups of low-light images of a single high-quality image, where each group of low-light images has a different low-light effect; by constructing a dimming module to generate a large number of paired samples of synthetic low-light images and high-quality images, the problem of scarcity of real paired data is effectively solved; the dimming module can simulate the imaging characteristics of different cameras under low-light conditions, so that the trained model has stronger generalization ability and is suitable for various types of camera equipment.

[0026] Step S12: Based on the dimming module, the high-quality image is converted into a synthetic image under low-light conditions to generate an artificial training data set; it is costly and difficult to obtain a large number of real high-quality images and corresponding low-light image pairs; by converting the high-quality image into a synthetic image under low-light conditions through the dimming module, a large number of simulated real low-light images can be generated; the demand for real paired data is significantly reduced, so that model training is no longer completely dependent on expensive data collection and annotation processes.

[0027] Fully supervised methods perform poorly when processing unseen data (such as images from different cameras); the synthetic data generated by the dimming module can simulate the imaging characteristics of different cameras under low-light conditions. Training with synthetic data can make the model better adapt to different types of camera equipment, thereby improving its generalization ability.

[0028] like Figure 2 As shown, Figure 2 A schematic diagram of the process of training a UNet model for a dimming module provided in the present application; Step S13: training an image enhancement model of a UNet (deep learning model) architecture in an unsupervised manner based on an artificial training dataset; the image enhancement model is trained based on the artificial training dataset to improve the brightness restoration level of low-light images; models that rely on a small amount of real data for training may not be well adapted to unseen data or scenes; artificially generated low-light images can cover a variety of lighting conditions, scene types, and noise characteristics, thereby increasing the diversity of training data, and the UNet model trained with diversified data exhibits stronger generalization capabilities when facing different scenes, devices, and lighting conditions.

[0029] Step S14: Implement image enhancement of low-light images based on the image enhancement model; the brightness value can be set during the image enhancement process; the brightness level of the output image is controlled based on the brightness value; during the image enhancement process, the user can specify a brightness value according to the needs As an additional conditional input, the brightness value is used to control the brightness level of the output image, allowing users to flexibly adjust the brightness of the enhanced image.

[0030] Users can flexibly adjust the brightness level of the output image according to specific needs without retraining the model. It is suitable for a variety of application scenarios such as night monitoring, photography post-processing, etc.; by precisely controlling the brightness value, users can obtain a more expected enhancement effect; the enhanced image is not only brighter, but also maintains natural colors and details, avoiding overexposure or distortion.

[0031] In summary, the semi-supervised low-light image enhancement method of this embodiment includes the following steps: constructing a dimming module based on a mixed intensity network, simulating the color distortion and dimming process of an image under low-light conditions based on the dimming module; converting a high-quality image into a synthetic image under low-light conditions based on the dimming module; training an image enhancement model of a UNet architecture in an unsupervised manner based on an artificial training data set; the image enhancement model improves the brightness restoration level of low-light images based on the training of the artificial training data set; realizing image enhancement of low-light images based on the image enhancement model; the brightness value can be set during the image enhancement process; and the brightness level of the output image is controlled based on the brightness value. By introducing the brightness value as a conditional input, image distortion caused by excessive enhancement is avoided, and the naturalness and applicability of the enhancement result are improved.

[0032] In one embodiment, constructing a dimming module includes the following steps:

[0033] The i-th channel of the reflectivity is calculated by channel-by-channel normalization. The formula is:

[0034] in, represents Hadamard (element-wise) partitioning, and denote the i-th channel of image I and reflectivity respectively; suppose we have an RGB image I with a size of 256×256 and want to generate a low-light image; the pixel value range of the input image I is [0, 255]; for each pixel, calculate the three channel values ​​of reflectivity R; for example, the RGB value of a pixel is [100, 150, 200], and the average value is (100+150+200) / 3=150; reflectivity =100 150=0.67, =150 150=1, =200 150=1.33; RGB images include 3 channels (red, green, blue). In this embodiment, is the reflectance of the red channel, is the reflectance of the green channel, is the reflectivity of the blue channel; in other embodiments, different types of images corresponding to different numbers of channels may be selected, and no limitation is made here.

[0035] like Figure 3 As shown, Figure 3 A schematic diagram of a process of creating a dark image by a dimming module provided in the present application according to an embodiment.

[0036] Through channel-by-channel normalization, the image is decomposed into illumination component and reflection component through Retinex (image enhancement theory and algorithm); the normalized reflectance can better represent the color characteristics of the object surface without being affected by the light intensity.

[0037] Adjust the reflection component R based on the mixture density network (MDN), which is based on the input reflection component R, the illumination component L, and the dimmed illumination component , output the adjusted reflection component ; The color distortion of the image under low light conditions is simulated to improve the authenticity of the generated low light image; In one embodiment, the reflection component R is [0.67, 1.0, 1.33], and after MDN adjustment, the adjusted reflection component is obtained =[0.5,0.8,1.1].

[0038] By introducing the mixed density network (MDN), the dimming module can flexibly adjust the reflection component to adapt to the imaging characteristics of different cameras under low-light conditions.

[0039] Based on the Hadamard product, the adjusted reflection component and the dimmed light component Combine to get the final dark image In one embodiment, the dimmed illumination component is set is [0.2, 0.3, 0.4], the final dark image =[0.5×0.2,0.8×0.3,1.1×0.4]=[0.1,0.24,0.44]; A realistic low-light image is generated by combining the adjusted reflection component with the dimmed illumination component.

[0040] Build a complete dimming module to generate realistic low-light images. The dimming module is implemented through channel-by-channel normalization, mixed density network (MDN), and Hadamard product operations. It can not only accurately simulate the image characteristics under low-light conditions, but also effectively reduce the dependence on real data, providing high-quality synthetic data for subsequent image enhancement model training.

[0041] In one embodiment, the mixed density network adjusts the reflection component R including the following steps: Model the reflectivity distortion of low-light images; regard the color distortion of low-light images as a probability distribution problem, and use MDN to model the color distortion of each pixel; the core idea of ​​MDN is to represent the output as a weighted combination of multiple Gaussian distributions, thereby capturing the complexity and uncertainty of color distortion.

[0042] Independently process the 5D representation of each pixel, which includes reflectance, illumination component L, and dimmed illumination component ; The reflectance includes three color channels (red, green, and blue); the 5-dimensional input of each pixel is processed independently to ensure that local characteristics can be accurately modeled; neural network layers are used to extract features and combined with MDN to generate probability distribution parameters for each color channel.

[0043] Output the mean, standard deviation and mixing coefficient of each color channel to describe the probability distribution of color distortion. The following describes the meaning of the output content in detail: Mean: Describes the central tendency of color distortion.

[0044] Standard Deviation: Describes the uncertainty or variance of color distortion.

[0045] Mixing coefficient: describes the weight of the Gaussian distribution in the overall distribution.

[0046] It should be clarified that in this embodiment, the present application does not limit the specific values ​​of the mean, standard deviation and mixing coefficient in the three different color channels of red, green and blue; the mean, standard deviation and mixing coefficient in different color channels can be the same or different, and there is no limitation on this.

[0047] By accurately modeling color distortion, the dimming module can generate more realistic low-light images; MDN can flexibly adjust the reflection component to adapt to the imaging characteristics of different cameras under low-light conditions; by independently processing the 5-D representation of each pixel, MDN can adapt to color distortion under various lighting conditions.

[0048] In one embodiment, the generation of an artificial training data set includes the following steps: Use high-quality images from public datasets as input; select publicly available computer vision datasets (such as ImageNet, COCO, etc.) to ensure that the image content is diverse and covers different scenes such as landscapes, people, and indoor environments.

[0049] The dimming module is called to convert the high-quality image into a synthetic image under low-light conditions; each high-quality image is input into the dimming module; and the output low-light image is a simulated version of the high-quality image under low-light conditions.

[0050] The high-quality images are paired with the corresponding low-light synthetic images to form training data pairs; multiple training data pairs constitute an artificial training dataset.

[0051] Through the above steps, an artificial training dataset can be efficiently generated for training low-light image enhancement models; this method not only reduces the dependence on real paired data, but also improves the diversity and quality of training data, thereby significantly improving the performance and generalization ability of the model.

[0052] In one embodiment, the dimming module converts the high-quality image into a synthetic image under low-light conditions and can set a dimming factor; the dimming factor is used to control the light level of the generated low-light image, and different dimming factor values ​​are set to generate low-light images with different light levels; the dimming factor It is a numerical parameter, usually ranging from 0< ≤1; when When it is close to 1, the generated low-light image is brighter; when When it is close to 0, the generated low-light image is very dark. Users can select different Values, such as =0.5, simulating moderately low light conditions; =0.2, simulating darker low-light conditions.

[0053] The dimming factor provides a flexible control method, which can generate low-light images of different degrees according to actual needs, so that the model can adapt to a wider range of low-light scenes and improve the diversity of training data; different dimming factor values ​​can generate low-light images under various lighting conditions, making the training data richer.

[0054] In one embodiment, the image enhancement model of the UNet architecture is fine-tuned based on real images to improve the adaptability of the UNet architecture to different cameras; artificially generated low-light images can cover a variety of scenes and lighting conditions, but they may not be able to fully reproduce the complex characteristics of real low-light images; fine-tuning based on real image pairs can make up for this deficiency and further improve the robustness and authenticity of the model.

[0055] In one embodiment, the UNet architecture includes a brightness setting module, which is used to receive a brightness value provided by a user. ; Set the brightness value based on the brightness setting module; the specific steps are as follows: Set the user-specified brightness value Embedded into the UNet architecture, the model can adjust the brightness of the output image as needed; The brightness value is transformed by linear transformation or nonlinear activation function Combined with the current feature map to control the brightness level of the output image.

[0056] In one embodiment, the UNet architecture receives as input a dimmed image, a dimmed image after histogram equalization, a color map, and lighting information; the following describes in detail what the input information represents: The dimmed image is the original low-light image; The dimmed image after histogram equalization is an image obtained by performing histogram equalization on the original low-light image, which enhances the contrast and highlights the details; Color mapping reflects the color distribution characteristics of the image and helps the model retain natural colors; The illumination information represents the distribution of illumination intensity in different regions of the image and is used to guide brightness adjustment.

[0057] Combined with the brightness value set by the brightness setting module, the residual image is output. The residual image is combined with the dimmed image to obtain the final enhanced image. The combination formula is as follows: To enhance the image, is the original low-light image, and U is the residual image.

[0058] UNet receives dimmed images, dimmed images after histogram equalization, color maps, and lighting information as input, and can understand the characteristics of low-light images from multiple angles; multi-dimensional information fusion helps the model to more accurately restore the brightness and details of the image while retaining natural colors; the introduction of the residual map enables the model to accurately adjust the brightness and details of the low-light image to avoid over-enhancement or distortion.

[0059] For the above embodiment, the present application provides a computer device, see Figure 4 , Figure 4 This is a schematic diagram of the structure of an embodiment of a computer device of the present application. The computer device includes a memory and a processor, wherein the memory and the processor are coupled to each other, the memory stores program data, and the processor is used to execute the program data to implement the steps of any embodiment of the above-mentioned semi-supervised low-light image enhancement method.

[0060] In this embodiment, the processor may also be referred to as a CPU (Central Processing Unit). The processor may be an integrated circuit chip having signal processing capabilities. The processor may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0061] The method of the above embodiment can be implemented in the form of a computer program, so the present application proposes a computer readable storage medium, see Figure 5 , Figure 5 This is a schematic diagram of the structure of an embodiment of a computer-readable storage medium of the present application. The computer-readable storage medium stores program data that can be run by a processor, and the program data can be executed by the processor to implement the steps of any embodiment of the above-mentioned semi-supervised low-light image enhancement method.

[0062] The computer-readable storage medium of this embodiment may be a medium that can store program data, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or may also be a server storing the program data, which may send the stored program data to other devices for execution, or may also execute the stored program data itself. In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only illustrative, for example, the division of the modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0063] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0064] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0065] The above description is only an implementation method of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A low-light image enhancement method based on semi-supervision, characterized in that: The steps include: Constructing a dimming module based on a hybrid intensity network, and simulating the color distortion and dimming process of an image under low light conditions based on the dimming module; Converting the high-quality image into a synthetic image under low-light conditions based on the dimming module to generate an artificial training dataset; Training an image enhancement model of a UNet architecture in an unsupervised manner based on the artificial training dataset; The image enhancement model is trained based on an artificial training data set to improve the brightness restoration level of low-light images; Implementing image enhancement of low-light images based on the image enhancement model; The brightness value can be set during image enhancement; A brightness level of an output image is controlled based on the brightness value.

2. The semi-supervised low-light image enhancement method according to claim 1, characterized in that: Constructing the dimming module includes the following steps: The i-th channel of the reflectivity is calculated by channel-by-channel normalization. The formula is: ； in, represents Hadamard (element-wise) partitioning, and denote the i-th channel of image I and reflectivity respectively; Adjust the reflection component R based on the mixture density network (MDN), which is based on the input reflection component R, the illumination component L, and the dimmed illumination component , output the adjusted reflection component ; Based on the Hadamard product, the adjusted reflection component With dimmed light component Combine to get the final dark image .

3. The semi-supervised low-light image enhancement method according to claim 2, characterized in that: The mixed density network adjusts the reflection component R and comprises the following steps: Modeling reflectivity distortion in low-light images; Independently process a 5-dimensional representation of each pixel, which includes reflectance, illumination component L, and dimmed illumination component ; The reflectivity includes three color channels; Outputs the mean, standard deviation, and mixing coefficient of each color channel, describing the probability distribution of color distortion.

4. The semi-supervised low-light image enhancement method according to claim 1, characterized in that: The generation of the artificial training data set comprises the following steps: Use high-quality images from public datasets as input; Call the dimming module to convert the high-quality image into a synthetic image under low-light conditions; The high-quality images are paired with the corresponding low-light synthetic images to form training data pairs; multiple training data pairs constitute an artificial training dataset.

5. The semi-supervised low-light image enhancement method according to claim 4, characterized in that: The dimming module converts the high-quality image into a synthetic image under low-light conditions and can set a dimming factor; the dimming factor is used to control the illumination level for generating the low-light image, and different dimming factor values ​​are set to generate low-light images with different illumination levels.

6. The semi-supervised low-light image enhancement method according to claim 1, characterized in that: It also includes fine-tuning the image enhancement model of the UNet architecture based on real images to improve the adaptability of the UNet architecture to different cameras.

7. The semi-supervised low-light image enhancement method according to claim 1, characterized in that: The UNet architecture includes a brightness setting module, and the setting of the brightness value is implemented based on the brightness setting module.

8. The semi-supervised low-light image enhancement method according to claim 7, characterized in that: The UNet architecture receives a dimmed image, a dimmed image after histogram equalization, a color map, and lighting information as input, and outputs a residual map in combination with the brightness value set by the brightness setting module. The residual map is combined with the dimmed image to obtain the final enhanced image.

9. An electronic device, characterized in that: The electronic device comprises: a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the semi-supervised low-light image enhancement method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program data, and the program data can be executed by a processor to implement the semi-supervised low-light image enhancement method according to any one of claims 1 to 8.

Images