Unsupervised exposure correction exposure deviation enhancement network model construction method and system

Abstract

The invention discloses an exposure deviation enhancement network model construction method and system for unsupervised exposure correction, relates to the technical field of computer vision, relieves the problem that an existing exposure correction technology strongly depends on a reference image, and solves the problems of complex color degradation, structure deficiency and the like in an image recovery process. Performing illuminance distribution statistics based on the well-exposed image, and making reference illuminance distribution; performing exposure correction based on the poor exposure image and the complementary exposure image and based on an exposure deviation enhancement network model to obtain corresponding enhanced correction images; designing a loss function based on the poor exposure image and the enhanced correction image thereof, the complementary exposure image and the enhanced correction image thereof, and the reference illumination distribution; and training the exposure deviation enhancement network model by using a loss function, and obtaining a final exposure deviation enhancement network model after training. The method is suitable for image processing in the field of outdoor scenes.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to the field of complex image exposure technology. Background Art

[0002] Image quality is affected by the exposure level of the device, and the quality usually significantly deteriorates. Overexposure will cause the image color to fade and colors to be missing, while underexposure will cause the image to be dim and the noise to increase significantly. This quality degradation seriously damages the visual experience and the accuracy of information transmission, and further affects downstream computer vision tasks, such as object detection, semantic segmentation, and autonomous driving, etc.

[0003] Although the camera has an automatic exposure mode that can determine the "ideal" exposure setting according to the scene brightness, uniformly adjusting the exposure across the entire image will still result in over-bright and over-dark areas. And these underexposed and overexposed areas will show obvious color distortion. With the progress of deep learning, many image enhancement methods have been proposed to correct images with different degrees of underexposure or overexposure. However, the current mainstream methods still have three levels of technical limitations: First, at the level of processing image types, the existing methods mainly target single-exposure images and have poor generalization ability for opposite exposures.

[0004] Second, at the level of lighting conditions, the existing methods introduce a unified network to handle the problems of underexposure and overexposure. However, they are based on uniform lighting conditions. When dealing with complex non-uniform exposure images, these methods often perform poorly and cannot correct the over-bright and over-dark areas in the same image simultaneously.

[0005] Third, at the level of data dependence, most of the existing methods rely on labor-intensive paired data annotation. High-quality paired reference images not only require a large amount of financial and human resources, but also often show obvious personal style biases. To alleviate this problem, many unsupervised methods have been proposed and have achieved remarkable results. However, these unsupervised methods mainly focus on underexposed (low-light) images and show limited enhancement ability for images with different exposures.

[0006] In summary, the existing image exposure correction technology has a strong dependence on reference images, and there are problems such as complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. Summary of the Invention

[0007] The present invention alleviates the strong dependence of the existing exposure correction technology on reference images and solves the problems such as complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. The present invention provides the following solutions: Solution 1. A method for constructing an exposure deviation enhancement network model for unsupervised exposure correction, comprising the following steps: Step S1: Obtain a training set, which includes underexposed images and properly exposed images in outdoor scenes; Step S2: Conduct illuminance distribution statistics on the properly exposed images in the training set to produce a reference illuminance distribution; Step S3: Perform a complementary operation on the underexposed images in the training set to obtain complementary exposure images; Step S4: Perform exposure correction on the underexposed images and their complementary exposure images respectively based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images; Step S5: Design a loss function based on the underexposed images and their enhanced corrected images in Step S4, the complementary exposure images and their enhanced corrected images in Step S4, and the reference illuminance distribution obtained in Step S2 , and use the loss function to train the exposure deviation enhancement network model, and finally obtain the exposure deviation enhancement network model after training.

[0008] Further, in an embodiment of the present invention, the process of performing exposure correction on the underexposed images or their complementary exposure images based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images is as follows: First, obtain the illuminance component of the underexposed image or its complementary exposure image based on the Retinex decomposition operation, and then obtain the initial exposure feature through Conv. The initial exposure feature passes through an illuminance-aware three-branch enhancement module to obtain the corrected exposure feature. After repeating the correction 4 times, based on the corrected exposure feature Concat, channel attention adaptive compression is performed on the underexposed image or its complementary exposure image to obtain the corresponding enhanced corrected image.

[0009] Further, in an embodiment of the present invention, the method of performing the Retinex decomposition operation on the underexposed image or its complementary exposure image is the same. Among them, the underexposed image is decomposed based on Retinex to obtain the illuminance component M and the reflection component R: ; ; where c ∈ {R, G, B} represents the RGB channel, and I is the underexposed image, .

[0010] Further, in an embodiment of the present invention, the method of obtaining the corrected exposure feature from the initial exposure feature through the illuminance-aware three-branch enhancement module is as follows: First, the initial exposure feature obtains the input feature based on three-branch feature extraction , the balanced feature and the complementary feature . Then, based on three illuminance perception functions and Conv, feature fusion is performed, and finally the corrected exposure feature is obtained.

[0011] Furthermore, in an embodiment of the present invention, the three-branch feature extraction is as follows: First, the initial exposure feature learns its input feature and its complementary feature based on a dual-branch convolution with shared weights designed by standard Conv and central difference convolution CDConv and its complementary feature . At the same time, the exposure feature estimates its balanced feature based on instance normalization IN .

[0012] Furthermore, in an embodiment of the present invention, the feature fusion based on three illuminance perception functions and Conv is as follows: First, the formulas of the three illuminance perception functions are: ; ; ; Among them, is the illuminance component of the input image. The input feature and the balanced feature are fused based on the function and Conv to obtain the fused feature . The complementary feature and the balanced feature are fused based on the function and Conv to obtain the fused feature . The fused features and are fused based on the function and Conv to obtain the enhanced feature .

[0013] Furthermore, in an embodiment of the present invention, in step S5, the loss function is: ; Among them, is the latent factor consistency loss, is the luminance distribution loss, is the structural similarity loss.

[0014] Solution 2: An unsupervised exposure correction system includes the following modules: Module 1: Used to obtain a training set, which includes poorly exposed images and well-exposed images in outdoor scenes; Module 2: Used to perform illuminance distribution statistics on well-exposed images in the training set to produce a reference illuminance distribution; Module 3: Used to perform complementary operations on poorly-exposed images in the training set to obtain complementary exposure images; Module 4: Used to perform exposure correction on the poorly-exposed images and their complementary exposure images respectively based on an exposure deviation enhancement network model to obtain corresponding enhanced corrected images; Module 5: Used to design a loss function L based on the poorly-exposed images and their enhanced corrected images in Module 4, the complementary exposure images and their enhanced corrected images in Module 4, and the reference illuminance distribution obtained in Module 2, and use the loss function L to train the exposure deviation enhancement network model. After training, the final exposure deviation enhancement network model is obtained.

[0015] Solution 3: An electronic device according to the present invention includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus; The memory is used to store a computer program; The processor is used to implement the steps of any of the above methods when executing the program stored on the memory.

[0016] Solution 4: A computer-readable storage medium according to the present invention, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0017] The present invention is a high-quality exposure deviation enhancement network model construction technology that is applicable to various complex exposures, effectively alleviating the problem of strong dependence on high-quality reference images in existing methods. The present invention has strong stability and adaptability, and is an unsupervised exposure correction method, effectively overcoming the problem of image degradation caused by poor exposure. The specific beneficial effects include: 1. The unsupervised exposure correction exposure bias enhancement network model obtained by the construction method of the present invention is used to implement exposure correction of images. The difference between using this model for unsupervised exposure correction and the prior art is that in the prior art, there are problems such as strong dependence on reference images in exposure correction technology, and complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. To solve the above technical problems, the present invention constructs an exposure bias enhancement network model. The exposure bias enhancement network model is composed of a combination of a three-branch feature extraction module and an illuminance perception feature fusion module and stacked. Finally, the final enhanced image is obtained by compressing channels based on channel attention, which can handle various complex exposure problems, correct the brightness and color of the image, avoid uneven brightness distribution and color distortion, and suppress noise, thereby optimizing detail preservation and calculation efficiency, and effectively solving the problems of complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process; 2. The unsupervised exposure correction exposure bias enhancement network model of the present invention only requires easily obtained images with poor exposure, greatly alleviating the strong dependence on reference images in existing exposure correction technologies; 3. The unsupervised exposure correction exposure bias enhancement network model of the present invention performs experiments on an image data set under poor exposure conditions and conducts comparative analysis with the prior art. The experimental data confirm that while correcting the image brightness, the present invention can effectively restore the original color and structure, reduce the noise level, and maintain the clarity of image details; 4. The method of the present invention is applicable to image processing in the field of outdoor scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and easy to understand from the following description of the embodiments in conjunction with the drawings, where: Figure 1 is the principle framework diagram of the exposure bias enhancement network model described in Embodiment 1; Figure 2 is the principle framework diagram of the three-branch enhancement module with illuminance perception described in Embodiment 1; Figure 3 is the comparison diagram of example pictures described in Embodiment 1, where (a) is the image with poor exposure before correction, and (b) is the image corrected by the method of the present invention; Figure 4 is the comparison diagram of example pictures described in Embodiment 1, where (a) is the image with poor exposure before correction, and (b) is the image corrected by the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will clearly and completely describe various embodiments of the present invention with reference to the accompanying drawings. The embodiments described by referring to the drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation to the present invention.

[0020] Embodiment 1. The method for constructing an exposure deviation enhancement network model for unsupervised exposure correction described in this embodiment includes the following steps: Step S1: Obtain a training set, which includes poorly exposed images and well-exposed images in outdoor scenes; Step S2: Perform illuminance distribution statistics on the well-exposed images in the training set to produce a reference illuminance distribution; Step S3: For the poorly exposed images in the training set, perform a complementary operation to obtain complementary exposure images; Step S4: Perform exposure correction on the poorly exposed images and their complementary exposure images respectively based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images; Step S5: Design a loss function based on the poorly exposed images and their enhanced corrected images in Step S4, the complementary exposure images and their enhanced corrected images in Step S4, and the reference illuminance distribution obtained in Step S2 , and use the loss function to train the exposure deviation enhancement network model. After training, the final exposure deviation enhancement network model is obtained.

[0021] In this embodiment, the well-exposed images in the training set in Step S1 are used to produce a reference illuminance distribution, and well-exposed images in a variety of outdoor scenes are preferably selected.

[0022] In this embodiment, the well-exposed images in Step S2 obtain their illuminance components based on Retinex decomposition operations, perform illuminance distribution statistics, and produce a reference illuminance distribution. The specific method is as follows: The Retinex decomposition operation is inspired by the biological vision model, and the equation is: ; where represents the RGB channels, is the illuminance component of the th well-exposed image, is the th well-exposed image.

[0023] Performing illuminance distribution statistics on the illuminance components of well-exposed images to produce a reference illuminance distribution The equation is: ; In the formula, For the reference illuminance distribution with good final exposure, it will guide the luminance distribution of the final corrected image to approach a good distribution. is the number of images with good exposure.

[0024] In this embodiment, complementary operations are performed on the images with poor exposure described in step S3 to obtain complementary exposure images, that is, for the images with poor exposure in the training set , the image Based on normalization, the values are within the range, and through complementary operations, complementary exposure images are obtained .

[0025] In this embodiment, the loss function described in step S5 trains the exposure deviation enhancement network model. Through the backpropagation algorithm, the entire network model is trained, and the parameters of the exposure deviation enhancement network model are updated. Among them, a judgment mechanism is implemented to determine whether a predetermined number of training iterations have been completed. When all training iterations are completed, a trained and optimized exposure deviation enhancement network model is obtained.

[0026] In this embodiment, preferably after the model training is completed, a test set is established to verify the exposure deviation enhancement network model of the present application. That is, a test set is established, and the test set is composed of images with poor exposure in outdoor scenes. The images with poor exposure in the test set are input into the exposure deviation enhancement network model trained in step S5. The exposure deviation enhancement network model will process the input images with poor exposure and output the final enhanced images.

[0027] This method in this embodiment is a technique for constructing an exposure deviation enhancement network model with high quality and applicable to various complex exposures. It effectively alleviates the problem of strong dependence on high-quality reference images in existing methods, overcomes the problem of image degradation caused by poor exposure, and guides high-quality exposure correction without supervision by statistically making a reference luminance distribution from a large number of images with good exposure in different scenes.

[0028] As Figure 3 and Figure 4 shown, it is a comparison diagram of the correction examples of the unsupervised exposure method described in this embodiment, that is, Figure 3 in, Figure 3 (a) is an image with poor exposure, and due to insufficient exposure, the picture is not clear. Figure 3 (b) is the image after exposure correction using the exposure deviation enhancement network model described in this embodiment. Compared with Figure 3 (a), it is significantly clearer.

[0029] Figure 4 in.Figure 4 (a) is a poor exposure image, which is unclear due to overexposure. Figure 4 (b) is the image after exposure correction using the exposure deviation enhancement network model described in this embodiment. Compared with Figure 4 (a), this image is significantly clearer.

[0030] Embodiment 2: This embodiment further limits the method for constructing the exposure deviation enhancement network model described in Embodiment 1. In this embodiment, the process of performing exposure correction based on the exposure deviation enhancement network model in step S4 to obtain the corresponding enhanced corrected image is as follows: First, obtain the illumination component of the poor exposure image or its complementary exposure image based on the Retinex decomposition operation, and then obtain the initial exposure feature through Conv. The initial exposure feature passes through the three-branch enhancement module for illumination perception to obtain the corrected exposure feature. After repeating the correction 4 times, based on the corrected exposure feature Concat, adaptively compress the channels of the poor exposure image or its complementary exposure image to obtain the corresponding enhanced corrected image.

[0031] In this embodiment, as Figure 1 shown, for the poor exposure image , obtain its illumination component based on the Retinex decomposition operation and its reflection component , and then obtain the initial exposure feature through Conv . The initial exposure feature passes through the three-branch enhancement module for illumination perception to obtain the corrected exposure feature . After repeating the correction 4 times, splice the exposure features obtained each time in the channel dimension, and adaptively compress the channels based on channel attention to obtain the final exposure-enhanced image .

[0032] In this embodiment, as Figure 1 shown, for the complementary exposure image CI, obtain its illumination component based on the Retinex decomposition operation , and then obtain the initial exposure feature through Conv . The initial exposure feature passes through the three-branch enhancement module for illumination perception to obtain the corrected exposure feature . After repeating the correction 4 times, splice the exposure features obtained each time in the channel dimension, and adaptively compress the channels based on channel attention to obtain the final exposure-enhanced image ECI.

[0033] This embodiment further defines step S4 and illustrates the solution of step S4. This method corrects the underexposed image based on the exposure deviation enhancement network model. The three-branch feature extraction module effectively obtains different exposure features. The illuminance perception feature fusion module finely fuses according to the exposure condition of the original image, and adaptively compresses the feature channels by using channel attention to obtain a high-quality enhanced image.

[0034] Embodiment 3. This embodiment further defines the method for constructing the exposure deviation enhancement network model described in Embodiment 2. In this embodiment, the method for the Retinex decomposition operation on the underexposed image or its complementary exposure image is the same. Among them, the underexposed image is decomposed based on Retinex to obtain the illuminance component M and the reflection component R: ; ; where c ∈ {R, G, B} represents the RGB channels, and I is the underexposed image. .

[0035] This embodiment further defines the method for constructing the exposure deviation enhancement network model described in Embodiment 2 and explains the Retinex decomposition operation. This Retinex decomposition operation is based on the Retinex theory. The Retinex theory is an image enhancement method that simulates the human visual system's ability to perceive light and color. The core idea is to decompose the brightness of the image into an illuminance component related to the ambient light and a reflection component related to the inherent characteristics of the image. Improving the image quality by adjusting these two components is inspired by the biological vision model.

[0036] Embodiment 4. This embodiment further defines the method for constructing the exposure deviation enhancement network model described in Embodiment 2. In this embodiment, the method for obtaining the corrected exposure features from the initial exposure features through the three-branch enhancement module with illuminance perception is as follows: First, the initial exposure features are used to obtain input features based on three-branch feature extraction , balanced features and complementary features , and then feature fusion is performed based on 3 illuminance perception functions and Conv, and finally the corrected exposure features are obtained.

[0037] This embodiment further defines the three-branch enhancement module for illumination perception, and describes the three-branch enhancement module for illumination perception. The three-branch enhancement module for illumination perception includes two parts: three-branch feature extraction and illumination perception feature fusion. This method obtains the corrected exposure feature, which prepares for adaptively compressing the feature channels using channel attention to obtain a high-quality enhanced image in the subsequent process.

[0038] Embodiment 5. This embodiment further defines the method for constructing the exposure bias enhancement network model described in Embodiment 4. In this embodiment, the three-branch feature extraction is as follows: First, the initial exposure feature learns its input feature and its complementary feature based on the dual-branch convolution with shared weights designed by the standard Conv and the central difference convolution CDConv. And its complementary feature At the same time, the exposure feature estimates its balanced feature based on the instance normalization IN. .

[0039] In this embodiment, as Figure 2 shown, the exposure feature and its complementary exposure feature activate and learn the exposure features with opposite characteristics based on the dual-branch convolution with shared weights, and obtain the input feature and the complementary feature . At the same time better preserves the inherent features such as the structural texture of the image based on the instance normalization (IN), and obtains the balanced feature . The input feature , the complementary feature and the balanced feature are fused based on 3 illumination perception fusion functions and Conv under the guidance of , and finally the corrected exposure feature is obtained.

[0040] The purpose of the dual-branch convolution is to enhance the model's perception ability of image details and textures, so as to extract the local details perception of different exposure features and improve the discriminability of the exposure feature expression. The dual-branch convolution consists of a standard Conv block and a central difference convolution block (CDConv) in parallel. The features output by channel splicing are fused by a 1x1 Conv and the number of channels is compressed and restored. The central difference convolution strengthens the ability to capture local structures such as edges and corners, and is more robust to noise and can suppress the influence of light changes or uniform noise.

[0041] This embodiment further defines the method for constructing the exposure bias enhancement network model, and describes the three-branch feature extraction. The three-branch feature extraction effectively obtains different exposure features.

[0042] Embodiment 6. In this embodiment, this embodiment further limits the method for constructing the exposure deviation enhancement network model described in Embodiment 4. In this embodiment, the feature fusion based on 3 illuminance perception functions and Conv is as follows: First, the formulas of the 3 illuminance perception functions are: ; ; ; Among them, is the input feature extracted by the three-branch feature extraction, is the complementary feature extracted by the three-branch feature extraction, is the balanced feature extracted by the three-branch feature extraction, is the illuminance component of the input image. The input feature and the balanced feature are fused based on the function and Conv to obtain the fused feature , and the complementary feature and the balanced feature are fused based on the function and Conv to obtain the fused feature . The fused features and are fused based on the function and Conv to obtain the enhanced feature .

[0043] This embodiment further limits the method for constructing the exposure deviation enhancement network model described in Embodiment 4, and illustrates the feature fusion of 3 illuminance perception functions and Conv. The enhanced feature obtained by this method improves the color and structure loss caused by overexposure, reduces the noise introduced by underexposure, and effectively reveals the color and texture details of the dim area.

[0044] Embodiment 7. This embodiment further limits the method for constructing the exposure deviation enhancement network model described in Embodiment 1. In this embodiment, in step S5, the loss function is: ; Among them, is the potential factor consistency loss, is the luminance distribution loss, is the structural similarity loss.

[0045] In this embodiment, specifically, two widely recognized factorization criteria, the Retinex theory and the Fourier transform, are used to formulate the potential factor consistency loss ; The Retinex theory holds that the reflection component of the image should remain unchanged during the enhancement process. The input image I and its enhanced image EI share the same reflection component. ECI is the enhanced image of CI, and 1 - ECI has the same reflection component as EI. I, EI, and 1 - ECI obtain their corresponding reflection components based on the Retinex decomposition operation. , and , the loss is calculated through the MSE loss function (mean squared error loss). , ensuring that the color tones and structural textures of the enhanced image are consistent, defined as: ; The Fourier transform is performed on images with different exposures to obtain the amplitude component and the phase component, which exhibit similar Fourier frequency characteristics. Specifically, the amplitude component mainly contains luminance information, while the phase component is more closely related to the structural details. Then, the input image I and its enhanced image EI should have similar phase components. ECI is the enhanced image of CI, and 1 - ECI has similar phase components as EI. I, EI, and 1 - ECI calculate their corresponding phase components based on the Fourier transform. , and , through the loss function (absolute value loss function), the loss is calculated. , further ensuring that the detailed textures and structural contours of the enhanced image are consistent, defined as: ; Therefore, the above loss functions are summarized as: ; Utilize the fact that images with good exposure have similar illumination distributions to guide the luminance distribution of the enhanced image. The illumination component of the enhanced image EI is obtained based on the Retinex decomposition operation. and its luminance distribution is calculated. , The KL divergence and the Wasserstein distance (EMD) are used to design the luminance distribution loss. , The KL divergence makes the two distributions overlap as much as possible, ensuring that the luminance distribution of the enhanced image is similar to the reference distribution. , combined with the EMD, can promote more adaptive exposure enhancement based on the intrinsic features of the image, defined as: ; Therefore, the above loss functions are summarized as: ; In the formula, is the total loss function, is the structural similarity loss.

[0046] This embodiment further defines the method for constructing an exposure bias enhancement network model, and describes the loss function. The potential factor consistency loss , ensures that the enhanced image is consistent with the input image in terms of basic tone, contour structure, and texture details, thereby preventing the enhanced output from showing excessive personal style bias and enhancing the generalization ability of the network. The brightness distribution loss , ensures that the enhanced image presents a high-quality and balanced brightness distribution. is the structural similarity loss, which ensures the structural quality of the enhanced image. In summary, this loss function can be better used for training the exposure bias enhancement network model of this application.

Claims

1. A method for constructing an exposure bias enhancement network model for unsupervised exposure correction, characterized in that It includes the following steps: Step S1: Obtain a training set, which includes underexposed images and well-exposed images in outdoor scenes; Step S2: Perform illuminance distribution statistics on the well-exposed images in the training set to create a reference illuminance distribution; Step S3: For the underexposed images in the training set, perform a complementary operation to obtain complementary exposure images; Step S4: Respectively perform exposure correction on the underexposed images and their complementary exposure images based on an exposure deviation enhancement network model to obtain corresponding enhanced corrected images; Step S5: Design a loss function based on the defective exposure image and its enhanced corrected image described in step S4, the complementary exposure image and its enhanced corrected image described in step S4, and the reference illuminance distribution obtained in step S2 , and use the loss function to train the exposure deviation enhancement network model, and finally obtain the final exposure deviation enhancement network model after training.

2. The method for constructing an exposure deviation enhancement network model according to claim 1, wherein The process of performing exposure correction on the underexposed images and their complementary exposure images based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images in Step S4 is as follows: First, based on the Retinex decomposition operation, obtain the illuminance component from the underexposed image or its complementary exposure image, and then obtain the initial exposure feature through Conv. The initial exposure feature passes through a three-branch enhancement module for illuminance perception to obtain the corrected exposure feature. After repeating the correction 4 times, based on the concatenation of the corrected exposure features, perform channel attention adaptive compression on the underexposed image or its complementary exposure image to obtain the corresponding enhanced corrected image.

3. The method for constructing an exposure deviation enhancement network model according to claim 2, wherein The method based on the Retinex decomposition operation for the poorly exposed image or its complementary exposed image is the same. Among them, the illumination component is obtained by performing Retinex decomposition on the poorly exposed image and the reflection component : ； ； Among them, represents the RGB channel, is an image with poor exposure, .

4. The method for constructing an exposure deviation enhancement network model according to claim 2, wherein The method for obtaining the corrected exposure feature from the initial exposure feature through the three-branch enhancement module for illuminance perception is as follows: First, the initial exposure feature is obtained based on the input feature through three-branch feature extraction , the balanced feature and the complementary feature . Then, feature fusion is performed based on three illuminance perception functions and Conv, and finally the corrected exposure feature is obtained.

5. The method for constructing an exposure deviation enhancement network model according to claim 4, wherein The method for three-branch feature extraction is as follows: First, the initial exposure feature learns its input feature based on a shared-weight double-branch convolution designed by standard Conv and central difference convolution CDConv in parallel and its complementary feature , and at the same time, the exposure feature estimates its balanced feature based on instance normalization IN .

6. The method for constructing an exposure deviation enhancement network model according to claim 4, wherein The method for feature fusion based on 3 illuminance perception functions and Conv is as follows: First, the 3 illuminance perception functions are respectively: ； ； ； Among them, is the illumination component of the input image, the input feature and the equalization feature are fused based on function and Conv to obtain the fused feature , the complementary feature and the equalization feature are fused based on function and Conv to obtain the fused feature , the fused feature and are fused based on function and Conv to obtain the enhanced feature .

7. The method for constructing an exposure deviation enhancement network model according to claim 1, characterized in that In the said step S5, the said loss function is as follows: ； Among them, is the potential factor consistency loss, is the brightness distribution loss, is the structural similarity loss.

8. An unsupervised exposure correction system, characterized in that, It includes the following modules: Module 1: Used to obtain a training set, which includes underexposed images and well-exposed images in outdoor scenes; Module 2: Used to perform illuminance distribution statistics on the well-exposed images in the training set to create a reference illuminance distribution; Module 3: Used to perform a complementary operation on the underexposed images in the training set to obtain complementary exposure images; Module 4: Used to respectively perform exposure correction on the underexposed images and their complementary exposure images based on an exposure deviation enhancement network model to obtain corresponding enhanced corrected images; Module Five: Design a loss function based on the defective exposure images and their enhanced corrected images described in Module Four, the complementary exposure images and their enhanced corrected images described in Module Four, and the reference illuminance distribution obtained in Module Two , and use the loss function to train the exposure deviation enhancement network model, and finally obtain the exposure deviation enhancement network model after training.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the steps of the method described in any one of claims 1-7 when executing the program stored on the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the steps of the method described in any one of claims 1-7.

Images