Overexposure and underexposure image enhancement method, device, equipment and medium

Abstract

The invention discloses an overexposure and underexposure image enhancement method, device, equipment and medium, and the method comprises the steps: enabling an image decomposition network to respectively receive a first illumination image and a second illumination image through employing two encoder-decoder networks sharing the weight, and extracting multi-scale illumination distribution information through multi-scale connection; the image reconstruction module outputs a corresponding reflectivity component and a brightness component, the image enhancement network adopts an image enhancement sub-network to adjust illumination distribution and suppress noise based on the reflectivity component and the brightness component, and the image reconstruction module multiplies the adjusted reflectivity component with an illumination image element by element and outputs an enhanced image. According to the method, the technical effects of effectively recovering image details, reducing noise interference and improving real-time performance under the overexposure or underexposure condition are reflected, the method is particularly suitable for complex environments such as power equipment monitoring, the contrast ratio of the enhanced image is better, artifacts are fewer, the calculation burden is reduced, and efficient batch processing prediction is supported.

Description

Technical Field

[0001] This application relates to the field of image enhancement technology, and more specifically, to a method, apparatus, device, and medium for enhancing overexposed and underexposed images. Background Technology

[0002] In recent years, with the widespread adoption of image acquisition equipment and the rapid development of deep learning vision algorithms, the requirements for automation, intelligence, and visualization in power systems have been increasing. Image processing technology, as one of the key supporting technologies for smart grids, has been widely and deeply applied in many fields such as power equipment detection, monitoring, and fault early warning. In the actual operation of power systems, power equipment is often located in complex and variable environments such as outdoor transmission lines and substations. These special working environments result in significant variations in lighting conditions, leading to underexposure or overexposure in captured images. Underexposure or overexposure can cause a series of problems, such as loss of image detail, reduced contrast, and increased noise interference. These problems can seriously interfere with the normal operation of image-based computer vision systems. For example, when detecting defects in power equipment, underexposure or overexposure images can make it difficult to accurately identify defects occurring during the operation of power facilities, thus threatening the safe and stable operation of the equipment.

[0003] In existing technologies, the enhancement of overexposed and underexposed images mainly employs two traditional methods: one is based on histogram equalization algorithms. This algorithm adjusts the image histogram to make the probability density of each gray level more uniform, thereby optimizing the image contrast and suppressing overexposure and underexposure. Histogram equalization, as a traditional and classic machine vision method, has the advantage of relatively simple computation, making it suitable for real-time processing. However, it also has significant drawbacks, namely, it easily leads to loss of image details and color distortion, especially when processing underexposed and overexposed areas, where the enhancement effect is often unsatisfactory.

[0004] The second approach is based on the classic Retinex theory. Retinex theory posits that an image can be decomposed into a reflectance map and an illumination map. By processing these two parts separately and then recombining them, the image's details and contrast can be enhanced, thus better recovering details in overexposed and underexposed areas. While this method has certain advantages, it also has several problems: First, the assumption that the reflectance component is the ideal result of the enhancement does not always hold true in practical applications. Considering various lighting characteristics, this may lead to unrealistic enhancement effects, such as loss of detail and color distortion. Second, the Retinex model typically ignores noise, allowing noise to be preserved or even amplified in the enhanced result. Finally, this method requires a sophisticated regularization algorithm; poor design, inaccurate priors, or inaccurate regularization can lead to artifacts and color casts in the enhanced image. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide a method, apparatus, device and medium for enhancing overexposed and underexposed images, which can be used to enhance overexposed and underexposed scenes in the field of power vision by using neural networks.

[0006] The objective of this application is achieved through the following technical solution: In a first aspect, this application proposes a method for enhancing overexposed and underexposed images. The method is applied to an overexposed and underexposed image enhancement device, which includes an image decomposition network, an image enhancement network, and an image reconstruction network. The method includes: The image decomposition network uses two encoder-decoder networks with shared weights to receive the first illumination image and the second illumination image respectively. It extracts multi-scale illumination distribution information through multi-scale connections and outputs the corresponding reflectance component and luminance component. The scene illumination intensity of the first illumination image is lower than that of the second illumination image. The image enhancement network is based on reflectance and luminance components, and uses an image enhancement sub-network to adjust the illumination distribution and suppress noise. The image reconstruction module multiplies the adjusted reflectance component element-wise with the illumination map to output an enhanced image.

[0007] In one possible implementation, the image decomposition network performs: Features of the input image are extracted using a 3×3 convolutional layer; The reflectance and illuminance components are mapped by a convolutional block consisting of five 3×3 convolutional layers and a ReLU activation function. Use the sigmoid function to constrain the component values ​​to the interval [0,1].

[0008] In one possible implementation, the image enhancement network includes: Three cascaded downsampling blocks, each consisting of a convolutional layer with a stride of 2 and a ReLU activation function; Three cascaded upsampling blocks, each consisting of nearest neighbor interpolation, a convolutional layer with a stride of 1, and a ReLU activation function; Residual connection between downsampling blocks and corresponding upsampling blocks.

[0009] In one possible implementation, the loss function of the overexposed / underexposed image enhancement device is: ,in To reconstruct the loss, For reflectivity uniformity loss, The coefficient for reflectivity consistency loss determines the importance of this constraint in the total loss, and is set to 0.001; To smooth out illuminance loss, This is the coefficient for illuminance smoothing loss, used to control the weight of the smoothness constraint on the total loss, and is set to 0.1.

[0010] In one possible implementation, the reflectivity uniformity loss is: ; Illuminance smoothing loss is: ; Reconstruction loss for: .

[0011] In one possible implementation, the method further includes: During training, the image decomposition network is fed in parallel with pairs of first and second illumination images, and the two branches share network weights. Only a single initial lighting image needs to be entered during testing.

[0012] Secondly, this application proposes an overexposed / underexposed image enhancement device, which includes an image decomposition network, an image enhancement network, and an image reconstruction network. The device comprises: An image decomposition network is used to receive a first illumination image and a second illumination image respectively using two encoder-decoder networks with shared weights. Multi-scale illumination distribution information is extracted through multi-scale connections, and the corresponding reflectivity component and luminance component are output. The scene illumination intensity of the first illumination image is lower than that of the second illumination image. An image enhancement network is used to adjust the illumination distribution and suppress noise based on the reflectivity and luminance components using an image enhancement subnetwork. The image reconstruction module is used to multiply the adjusted reflectance component element-wise with the illumination map to output an enhanced image.

[0013] Thirdly, this application also proposes a computer device comprising a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the overexposure and underexposure image enhancement method as described in any of the first aspects.

[0014] Fourthly, this application also proposes a computer-readable storage medium storing a computer program that is loaded and executed by a processor to implement the overexposure and underexposure image enhancement method as described in any of the first aspects.

[0015] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0016] This application discloses a method, apparatus, device, and medium for enhancing overexposed and underexposed images. First, the image decomposition network uses two encoder-decoder networks with shared weights to receive a first illumination image and a second illumination image, respectively. Multi-scale illumination distribution information is extracted through multi-scale connections, and the corresponding reflectance and luminance components are output. Based on the reflectance and luminance components, the image enhancement network uses an image enhancement sub-network to adjust the illumination distribution and suppress noise. The image reconstruction module multiplies the adjusted reflectance component with the illumination image element by element to output the enhanced image. This method demonstrates the technical effects of effectively restoring image details, reducing noise interference, and improving real-time performance under overexposed or underexposed conditions. It is particularly suitable for complex environments such as power equipment monitoring. The enhanced image has better contrast, fewer artifacts, and reduced computational burden, supporting efficient batch processing prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of an overexposed / underexposed image enhancement method according to an embodiment of this application is shown.

[0019] Figure 2 A schematic diagram of the structure of the overexposed and underexposed image enhancement device proposed in an embodiment of this application is shown. Detailed Implementation

[0020] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0021] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In existing technologies, histogram equalization (HOG) methods adjust the probability density of each gray level in an image by modifying its histogram (HOG) to make the pixel value distribution more uniform, thereby optimizing image contrast and suppressing overexposure and underexposure. Histogram equalization is a traditional and classic machine vision method with the advantage of computational simplicity, making it suitable for real-time processing. However, this method also has the following drawbacks: it easily leads to loss of detail and color distortion, especially when dealing with underexposed and overexposed areas. Methods based on typical Retinex theory consider an image to be decomposed into a reflectance map and an illumination map. By separating these two parts, they can be processed separately and then recombined to enhance image detail and contrast, achieving better recovery of details in overexposed and underexposed areas. However, this method also has the following problems: 1) The assumption that the reflectance component is considered an ideal result of the enhancement does not always hold, especially considering various lighting characteristics, which may lead to unrealistic enhancements such as loss of detail and color distortion; 2) Noise is usually ignored in the Retinex model, so noise is either preserved or amplified in the enhancement result. 3) This method requires the design of an ingenious regularization algorithm. Poorly designed or inaccurate priors or regularization may lead to artifacts and color distortion in the enhanced results.

[0023] Therefore, to address the aforementioned technical problems, this application proposes a method, apparatus, device, and medium for enhancing overexposed and underexposed images. It employs an image decomposition network with two weighted encoder-decoder networks to process a first and second illumination image respectively. Multi-scale connections are used to extract multi-scale illumination distribution information, outputting reflectance and luminance components. An image enhancement network then adjusts the illumination distribution and suppresses noise based on these components. Finally, an image reconstruction module multiplies the adjusted reflectance component element-wise with the illumination image to output the enhanced image. This demonstrates the technical effectiveness of effectively restoring image details, reducing noise interference, and improving real-time performance under overexposed or underexposed conditions. It is particularly suitable for complex environments such as power equipment monitoring, producing enhanced images with better contrast, fewer artifacts, and reduced computational burden, supporting efficient batch processing prediction.

[0024] Please refer to Figure 1 , Figure 1 This illustration shows a flowchart of an overexposed / underexposed image enhancement method according to an embodiment of this application. The method is applied to an overexposed / underexposed image enhancement device, which includes an image decomposition network, an image enhancement network, and an image reconstruction network. The method includes: Step S1: The image decomposition network uses two encoder-decoder networks with shared weights to receive the first illumination image and the second illumination image, respectively. Multi-scale illumination distribution information is extracted through multi-scale connections, and the corresponding reflectance and luminance components are output. The scene illumination intensity in the first illumination image is lower than that in the second illumination image; Image data of the same object scene under good lighting and poor lighting conditions were collected using an image acquisition card and a visible light camera, respectively. After manual inspection and pairing, a training dataset was obtained. The first and second lighting images were then extracted from this dataset.

[0025] The image decomposition network employs two weighted encoder-decoder networks, receiving images with higher illumination and lower illumination respectively. Through multi-scale connections, it extracts multi-scale illumination distribution information and outputs the corresponding reflectance and luminance components. The network module first downsamples the input image to obtain large-scale illumination distribution information, thereby grasping the global illumination characteristics of the image. Subsequently, it uses upsampling blocks to reconstruct the local illumination distribution, thereby recovering the detailed information in the image. To finely adjust the illumination from a hierarchical perspective—that is, to maintain global illumination consistency while making corresponding adjustments for different local illumination distributions—the network introduces a multi-scale connection mechanism.

[0026] Image decomposition network execution: Features of the input image are extracted using a 3×3 convolutional layer; The reflectance and illuminance components are mapped by a convolutional block consisting of five 3×3 convolutional layers and a ReLU activation function. Use the sigmoid function to constrain the component values ​​to the interval [0,1].

[0027] First, a 3×3 convolutional layer is used to extract features from the input low-light and normal-light images, obtaining basic image feature information through convolution operations. Next, the extracted features are sequentially passed through five convolutional blocks consisting of 3×3 convolutional layers and a ReLU activation function. The convolutional operations further extract and transform features, while the ReLU activation function introduces non-linearity, enabling the network to learn more complex feature representations, thus mapping the RGB image into reflectance and luminance components. Subsequently, a 3×3 convolutional layer projects the reflectance and illumination from the feature space to better separate and adjust the image's reflectance and illumination information. Finally, the sigmoid function constrains the values ​​of reflectance and illumination within the [0,1] interval, ensuring that the values ​​of the reflectance and luminance components are within a reasonable range.

[0028] Step S2: The image enhancement network adjusts the illumination distribution and suppresses noise based on the reflectivity and luminance components using an image enhancement sub-network.

[0029] In the illumination adjustment step, the image enhancement network uses the reflectance and luminance components obtained from the image decomposition network as input. This image enhancement sub-network, based on an encoder-decoder architecture, processes the input reflectance and luminance components to effectively adjust the image illumination distribution and performs noise reduction on overexposed and underexposed areas, thereby improving image quality and enhancing image details and contrast.

[0030] Image enhancement networks include: Three cascaded downsampling blocks, each consisting of a convolutional layer with a stride of 2 and a ReLU activation function; Three cascaded upsampling blocks, each consisting of nearest neighbor interpolation, a convolutional layer with a stride of 1, and a ReLU activation function; Residual connection between downsampling blocks and corresponding upsampling blocks.

[0031] The image enhancement network takes a low-light image as input and performs a downsampling operation on the input image to make the network more focused on lighting characteristics. The downsampling operation consists of three downsampling blocks, each consisting of a convolutional layer with a stride of 2 and a ReLU activation function. After downsampling, the network performs an upsampling layer: having acquired a large amount of lighting information, the upsampling operation amplifies the lighting. The upsampling step consists of three upsampling blocks, each consisting of a nearest neighbor interpolation operation, a convolutional layer with a stride of 1, and a ReLU+ activation function.

[0032] After upsampling, the module imports the summation of elements from the downsampled block into the corresponding mirror upsampled block, thus forcing the network to learn residuals. The output of the residuals is then used to adjust the size of these features at different scales using nearest neighbor interpolation to the final scale, and output as feature maps that match the feature size. Then, cascaded features are restored to the original number of channels using 1×1 convolutional layers, and finally, the illumination map is reconstructed using 3×3 convolutional layers and output through the output layer.

[0033] Step S3: The image reconstruction module multiplies the adjusted reflectance component element-wise with the illumination map to output the enhanced image.

[0034] The image reconstruction module receives the reflectance component and illumination map adjusted by the image enhancement network. Using the Retinex model, it multiplies corresponding elements of the two to reconstruct the enhanced image. This operation effectively fuses the detail information in the reflectance component and the illumination information in the illumination map, significantly improving the output image in terms of illumination, contrast, and detail. This enables enhancement processing of overexposed and underexposed images, providing clearer and more accurate image data for tasks such as power equipment inspection.

[0035] The loss function of the overexposed / underexposed image enhancement device is: ,in To reconstruct the loss, For reflectivity uniformity loss, This is for illuminance smoothing loss.

[0036] Reconstruction loss for: This loss function measures the difference between the reconstructed image and the original image by comparing the result reconstructed based on the reflectance component and the illumination component with the original image, thereby ensuring that the reconstructed image is as close to the original image as possible.

[0037] The reflectivity uniformity loss is: According to Retinex theory, the reflectance components under low-light and normal-light conditions should be kept as consistent as possible. Therefore, this loss function is used to constrain the reflectance components of low-light images. Reflectance component of normal illumination image The differences between them are to ensure consistency between the two.

[0038] Illuminance smoothing loss is: This loss function aims to minimize the gradient of the illumination component at texture details, thereby achieving noise reduction for overexposed and underexposed areas. This is achieved by incorporating the gradient of the illumination component. gradient of reflectivity component It can enhance images while retaining more detail, thus improving the overall image quality.

[0039] Overexposed and underexposed image enhancement methods also include: During training, the image decomposition network is fed in parallel with pairs of first and second illumination images, and the two branches share network weights. Only a single initial lighting image needs to be entered during testing.

[0040] During training, pairs of first and second illumination images are input in parallel to the image decomposition network. The first illumination image is a low-light image, and the second illumination image is a normal-light image. The image decomposition network contains two structurally identical branches that share network weights. These branches extract features and decompose the input images, obtaining the reflectance and luminance components corresponding to the low-light and normal-light images, respectively. During testing, only a single first illumination image (the low-light image) needs to be input to the image decomposition network. The network can then decompose the input low-light image based on the weight parameters learned during training, obtaining its reflectance and luminance components. Throughout the training process, it is not necessary to provide the true values ​​of reflectance and illumination. Instead, reflectance consistency and illumination smoothness are embedded as loss functions into the network to guide it in learning effective image decomposition features, thereby achieving accurate image decomposition.

[0041] The following describes a possible implementation of an overexposed / underexposed image enhancement device, which performs the various execution steps and corresponding technical effects of the overexposed / underexposed image method shown in the above embodiments and possible implementations. The overexposed / underexposed image enhancement device includes an image decomposition network, an image enhancement network, and an image reconstruction network. The device includes: An image decomposition network is used to receive a first illumination image and a second illumination image respectively using two encoder-decoder networks with shared weights. Multi-scale illumination distribution information is extracted through multi-scale connections, and the corresponding reflectivity component and luminance component are output. The scene illumination intensity of the first illumination image is lower than that of the second illumination image. An image enhancement network is used to adjust the illumination distribution and suppress noise based on the reflectivity and luminance components using an image enhancement subnetwork. The image reconstruction module is used to multiply the adjusted reflectance component element-wise with the illumination map to output an enhanced image.

[0042] Figure 2 This paper illustrates a schematic diagram of the overexposed and underexposed image enhancement device proposed in an embodiment of this application. In the image decomposition module, the network uses two image decomposition network branches with shared weights to decompose the input normal illumination image. and low-light images The processing proceeds. Each branch contains multiple convolutional layers (Conv), convolutional ReLU activation layers (Conv+ReLU), and a sigmoid layer. Through these layers, the network decomposes the input image into reflectance components (…). , ) and brightness component ( , The two branches share weights to ensure consistent feature extraction and decomposition of images under different lighting conditions.

[0043] In the illumination adjustment module, the image enhancement network processes the reflectance component of the low-illumination image from the image decomposition network. and brightness component Further processing is then performed. This module contains multiple convolutional ReLU activation layers (Conv+ReLU), as well as resize, concat operations, and 1×1 convolutional layers. Through these operations, the network denoises and adjusts the illumination of the luminance component to generate an adjusted luminance component. In addition, this module also includes skip connections (shown by blue arrows) to help the network learn residuals and effectively utilize feature information.

[0044] In the reconstruction module, the enhanced luminance component With reflectivity component Element-wise multiplication generates the final enhanced image. This Retinex model-based element-wise multiplication operation effectively fuses reflectivity and brightness information, thereby enhancing overexposed and underexposed images.

[0045] During training, the network receives pairs of low-light and normal-light images as input, and learns image decomposition and enhancement features by optimizing loss functions such as reflectivity consistency and illumination smoothness. During testing, the network only requires the low-light image as input to output the enhanced image. By stacking images from multiple power scene monitoring cameras in batches and feeding them into the model for prediction, the enhancement results can be output more efficiently, improving the accuracy and efficiency of tasks such as power equipment detection.

[0046] Compared with the prior art, the embodiments of this application have the following beneficial effects: First, deep learning networks based on Retinex theory exhibit significant advantages in image enhancement. Retinex theory provides physical interpretability for network design, enabling the network to better understand the lighting characteristics of images when processing overexposed and underexposed images. This theoretical foundation allows the network to more effectively restore image details and contrast in various application scenarios, especially in power scene image enhancement, demonstrating superior enhancement results compared to traditional algorithms and end-to-end deep learning methods.

[0047] Secondly, in addition to following Retinex theory to divide illumination into intensity and reflectance, a dedicated sub-network was designed. The image decomposition module can extract the light-independent reflectance component and the structure-aware smooth illumination component from the input image, while the image enhancement module focuses on adjusting the illuminance map to enhance low-light images. This modular network design not only improves enhancement performance but also makes the network more focused on specific tasks, thereby improving overall efficiency.

[0048] Third, through specialized sub-network design, this solution excels in reducing computational burden. The design of the image decomposition and image enhancement modules enables the network to operate with fewer computational resources while ensuring real-time computation. This is particularly important for applications requiring real-time processing, such as power equipment monitoring, ensuring that image enhancement results can be quickly fed back to the monitoring system, thereby improving the efficiency of equipment detection and fault early warning.

[0049] This preferred embodiment provides a computer device that can implement the steps of any embodiment of the overexposed and underexposed image enhancement method provided in this application. Therefore, it can achieve the beneficial effects of the overexposed and underexposed image enhancement method provided in this application. For details, please refer to the previous embodiments, which will not be repeated here.

[0050] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of this application provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the overexposed / underexposed image enhancement method provided in this application.

[0051] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0052] Since the instructions stored in the storage medium can execute the steps in any of the overexposed and underexposed image enhancement method embodiments provided in this application, the beneficial effects that any of the overexposed and underexposed image enhancement methods provided in this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0053] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for enhancing overexposed and underexposed images, characterized in that, The method is applied to an overexposed / underexposed image enhancement device, the device comprising an image decomposition network, an image enhancement network, and an image reconstruction network, and the method includes: The image decomposition network uses two encoder-decoder networks with shared weights to receive the first illumination image and the second illumination image respectively. It extracts multi-scale illumination distribution information through multi-scale connections and outputs the corresponding reflectance component and luminance component. The scene illumination intensity of the first illumination image is lower than that of the second illumination image. The image enhancement network is based on reflectance and luminance components, and uses an image enhancement sub-network to adjust the illumination distribution and suppress noise. The image reconstruction module multiplies the adjusted reflectance component element-wise with the illumination map to output an enhanced image.

2. The image enhancement method for overexposure and underexposure as described in claim 1, characterized in that, The image decomposition network performs the following: Features of the input image are extracted using a 3×3 convolutional layer; The reflectance and illuminance components are mapped by a convolutional block consisting of five 3×3 convolutional layers and a ReLU activation function. Use the sigmoid function to constrain the component values ​​to the interval [0,1].

3. The image enhancement method for overexposure and underexposure as described in claim 1, characterized in that, The image enhancement network includes: Three cascaded downsampling blocks, each consisting of a convolutional layer with a stride of 2 and a ReLU activation function; Three cascaded upsampling blocks, each consisting of nearest neighbor interpolation, a convolutional layer with a stride of 1, and a ReLU activation function; Residual connection between downsampling blocks and corresponding upsampling blocks.

4. The image enhancement method for overexposure and underexposure as described in claim 1, characterized in that, The loss function of the overexposed / underexposed image enhancement device is: ,in To reconstruct the loss, For reflectivity uniformity loss, The coefficient for reflectivity consistency loss determines the importance of this constraint in the total loss, and is set to 0.001; To smooth out illuminance loss, This is the coefficient for illuminance smoothing loss, used to control the weight of the smoothness constraint on the total loss, and is set to 0.

1.

5. The image enhancement method for overexposure and underexposure as described in claim 4, characterized in that, The reflectivity uniformity loss is: , For low-light images, reflectivity The reflectance of a normal light image; Illuminance smoothing loss is: , It is a lighting diagram The gradient reflects the degree of change in light intensity. It is a reflectance map The gradient; Reconstruction loss for: , This is a reflectance map. This is a lighting diagram. The original input image, It is element-wise multiplication. It is an L1 norm.

6. The image enhancement method for overexposure and underexposure as described in claim 1, characterized in that, The method further includes: During training, the image decomposition network is fed in parallel with pairs of first and second illumination images, and the two branches share network weights. Only a single initial lighting image needs to be entered during testing.

7. An image enhancement device for overexposed and underexposed images, characterized in that, The overexposed / underexposed image enhancement device includes an image decomposition network, an image enhancement network, and an image reconstruction network. The device comprises: An image decomposition network is used to receive a first illumination image and a second illumination image respectively using two encoder-decoder networks with shared weights. Multi-scale illumination distribution information is extracted through multi-scale connections, and the corresponding reflectivity component and luminance component are output. The scene illumination intensity of the first illumination image is lower than that of the second illumination image. An image enhancement network is used to adjust the illumination distribution and suppress noise based on the reflectivity and luminance components using an image enhancement subnetwork. The image reconstruction module is used to multiply the adjusted reflectance component element-wise with the illumination map to output an enhanced image.

8. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the overexposed and underexposed image enhancement method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the overexposed and underexposed image enhancement method as described in any one of claims 1-6.

Images