Exposure bias enhancement network model construction method and system for unsupervised exposure correction

Abstract

The invention relates to the field of computer vision technology, which alleviates the problem of strong dependence of existing exposure correction technology on reference images and solves the problems of complex color degradation, structure loss and so on in the image restoration process. Based on the well-exposed image, the illumination distribution statistics are performed to produce a reference illumination distribution; based on the poorly exposed image and its complementary exposed image, exposure correction is performed based on the exposure deviation enhancement network model to obtain the corresponding enhanced corrected image; based on the poorly exposed image and its enhanced corrected image, the complementary exposed image and its enhanced corrected image, and the reference illumination distribution, a loss function is designed; the exposure deviation enhancement network model is trained using the loss function to obtain the final exposure deviation enhancement network model after training. The method of the present invention is suitable for image processing in outdoor scenes.

Description

Technical Field

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

[0002] Image quality is affected by the device's exposure level, often resulting in significant degradation. Overexposure results in faded and missing colors, while underexposure results in dark images with significant noise. This degradation severely impairs the visual experience and the accuracy of information transmission, impacting downstream computer vision tasks such as object detection, semantic segmentation, and autonomous driving.

[0003] Although cameras have automatic exposure modes that determine the "ideal" exposure setting based on scene brightness, adjusting exposure uniformly across the entire image can still result in overly bright and dark areas. These underexposed and overexposed areas exhibit noticeable tonal distortion. With advances in deep learning, many image enhancement methods have been proposed to correct images that are underexposed or overexposed to varying degrees. However, current mainstream methods still have technical limitations at three levels:

[0004] First, at the image type level, existing methods mainly target single-exposure images, but have poor generalization capabilities for images with opposite exposures.

[0005] Second, at the lighting level, existing methods introduce a unified network to handle 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 are unable to simultaneously correct overly bright and dark areas in the same image.

[0006] Third, regarding data dependency, existing methods mostly rely on labor-intensive paired data annotation. Producing high-quality paired reference images not only requires significant financial and human resources but also often exhibits significant bias in individual style. To alleviate this problem, numerous unsupervised methods have been proposed, achieving significant results. However, these unsupervised methods primarily focus on underexposed (low-light) images and exhibit limited ability to enhance images with varying exposures.

[0007] In summary, existing image exposure correction technology has a strong dependence on reference images, and there are complex problems such as color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. Summary of the Invention

[0008] This invention alleviates the existing exposure correction technology's strong reliance on reference images and solves problems such as complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. The invention provides the following solutions:

[0009] Solution 1: A method for constructing an exposure bias enhancement network model for unsupervised exposure correction, including the following steps:

[0010] Step S1: obtaining a training set, where the training set includes poorly exposed images and well-exposed images in outdoor scenes;

[0011] Step S2: performing illumination distribution statistics on the well-exposed images in the training set to produce a benchmark illumination distribution;

[0012] Step S3: performing a complementary operation on the poorly exposed images in the training set to obtain complementary exposure images;

[0013] Step S4: performing exposure correction on the poorly exposed image and its complementary exposed image based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images; specifically,

[0014] First, the illumination component of the poorly exposed image or its complementary exposed image is obtained based on the Retinex decomposition operation. Then, the initial exposure features are obtained through Conv. The initial exposure features are then subjected to the illumination-aware three-branch enhancement module to obtain the corrected exposure features (Concat). After repeating the correction four times, the iteratively corrected exposure features (Concat) are obtained.

[0015] Then, based on the iteratively corrected exposure feature Concat, a corresponding enhanced corrected image is obtained;

[0016] The process of obtaining the corrected exposure characteristics from the initial exposure characteristics through the illumination perception three-branch enhancement module is as follows:

[0017] First, the initial exposure features are extracted based on three-branch features to obtain input features. , balanced characteristics and complementary features , then perform feature fusion based on the three illumination perception functions and Conv, and finally obtain the corrected exposure feature Concat;

[0018] Step S5: Design a loss function based on the poorly exposed image and the enhanced corrected image described in step S4, the complementary exposed image and the enhanced corrected image described in step S4, and the reference illumination distribution obtained in step S2. , using the loss function The exposure bias enhancement network model is trained to obtain a final exposure bias enhancement network model after training.

[0019] Furthermore, in one embodiment of the present invention, the process of obtaining the corresponding enhanced corrected image based on the iteratively corrected exposure feature Concat in step S4 is:

[0020] Based on the iteratively corrected exposure feature Concat, channel attention adaptive compression is performed on the poorly exposed image or its complementary exposed image to obtain a corresponding enhanced corrected image.

[0021] Furthermore, in one embodiment of the present invention, the poorly exposed image or its complementary exposed image is subjected to the same method of Retinex decomposition operation, wherein the illumination component is obtained by Retinex decomposition of the poorly exposed image. and the reflected component :

[0022]

[0023] ;

[0024] in, represents RGB channels, is a poorly exposed image. .

[0025] Furthermore, in one embodiment of the present invention, the three-branch feature extraction is:

[0026] First, the initial exposure feature is learned based on the shared weight dual-branch convolution designed in parallel by standard Conv and center difference convolution CDConv. and its complementary characteristics At the same time, the exposure feature estimates its balanced feature based on instance normalization IN .

[0027] Furthermore, in one embodiment of the present invention, the feature fusion based on the three illumination perception functions and Conv is as follows:

[0028] First, the three illumination perception function formulas are:

[0029] ;

[0030] ;

[0031] ;

[0032] in, is the illumination component of the input image, the input features and the equalization features are based on Function and Conv are fused to obtain fusion features , complementary features and balanced features are based on Function and Conv are fused to obtain fusion features , fusion features and based on Function and Conv fusion to obtain enhanced features .

[0033] Furthermore, in one embodiment of the present invention, in step S5, the loss function is:

[0034] ;

[0035] in, is the latent factor consistency loss, is the brightness distribution loss, It is the structural similarity loss.

[0036] Solution 2: An unsupervised exposure correction system, including the following modules:

[0037] Module 1: used to obtain a training set, which includes poorly exposed images and well-exposed images of outdoor scenes;

[0038] Module 2: Used to calculate the illumination distribution statistics of well-exposed images in the training set and create a benchmark illumination distribution;

[0039] Module 3: Used to perform complementary operations on poorly exposed images in the training set to obtain complementary exposure images;

[0040] Module 4: for performing exposure correction on the poorly exposed image and its complementary exposed image based on an exposure deviation enhancement network model to obtain corresponding enhanced corrected images;

[0041] The module 4 further includes:

[0042] Submodule 1: Used to obtain the illumination component of the poorly exposed image or its complementary exposed image based on the Retinex decomposition operation, then obtain the initial exposure feature through Conv, and obtain the corrected exposure feature Concat after the initial exposure feature is processed through the illumination-aware three-branch enhancement module. After repeating the correction four times, the iteratively corrected exposure feature Concat is obtained.

[0043] Submodule 2: for obtaining a corresponding enhanced corrected image based on the iteratively corrected exposure feature Concat;

[0044] The submodule 1 further includes a unit for extracting input features based on the three-branch feature extraction of the initial exposure features. , balanced characteristics and complementary features , then perform feature fusion based on the three illumination perception functions and Conv, and finally obtain the corrected exposure feature Concat;

[0045] Module 5: Designing a loss function based on the poorly exposed image and its enhanced corrected image described in Module 4, the complementary exposed image and its enhanced corrected image described in Module 4, and the reference illumination distribution obtained in Module 2 , using the loss function The exposure bias enhancement network model is trained to obtain a final exposure bias enhancement network model after training.

[0046] Solution 3: An electronic device according to the present invention comprises a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0047] Memory for storing computer programs;

[0048] The processor is configured to implement the steps of any of the above methods when executing the program stored in the memory.

[0049] Solution 4: A computer-readable storage medium according to the present invention stores a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0050] This invention is a high-quality exposure deviation enhancement network model construction technology that is universally applicable to various complex exposures. It effectively alleviates the existing methods' strong dependence on high-quality reference images. The invention has strong stability and adaptability, and the unsupervised exposure correction method effectively overcomes the image degradation problem caused by poor exposure. Specific beneficial effects include:

[0051] 1. The exposure deviation enhancement network model for unsupervised exposure correction obtained by the construction method described in the present invention is used to realize exposure correction of images. The difference between using this model for unsupervised exposure correction and the prior art is that, in the prior art, the exposure correction technology has a strong dependence on the reference image, and there are problems such as complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process. In order to solve the above technical problems, the present invention constructs an exposure deviation enhancement network model. The exposure deviation enhancement network model is composed of a combination and stacking of a three-branch feature extraction module and an illumination perception feature fusion module, and finally obtains the final enhanced image based on the channel attention compression channel. It 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 retention and computational efficiency, and effectively solving the problems of complex color degradation, structure loss, dim brightness, and excessive noise in the image restoration process;

[0052] 2. The exposure deviation enhancement network model for unsupervised exposure correction described in the present invention only requires easily accessible poorly exposed images, which greatly alleviates the problem of existing exposure correction technology's strong dependence on reference images.

[0053] 3. The exposure deviation enhancement network model for unsupervised exposure correction described in the present invention was tested on an image dataset under poor exposure conditions and compared with the existing technology. The experimental data confirmed that the present invention can effectively restore the original color and structure, reduce the noise level and maintain the clarity of image details while correcting the image brightness.

[0054] The method of the present invention is applicable to image processing in the field of outdoor scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0056] Figure 1 This is a principle framework diagram of the exposure bias enhancement network model described in Implementation Method 1.

[0057] Figure 2 This is a principle framework diagram of the three-branch enhancement module for illumination perception described in Implementation Method 1.

[0058] Figure 3 This is a comparison diagram of example pictures described in the first embodiment, wherein (a) is a poorly exposed image before correction, and (b) is an image corrected using the method described in the present invention.

[0059] Figure 4This is a comparison diagram of example pictures described in the first embodiment, wherein (a) is a poorly exposed image before correction, and (b) is an image corrected using the method described in the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe various embodiments of the present invention in conjunction with the accompanying drawings. The embodiments described with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0061] Implementation method 1: The method for constructing an exposure bias enhancement network model for unsupervised exposure correction described in this implementation method includes the following steps:

[0062] Step S1: obtaining a training set, where the training set includes poorly exposed images and well-exposed images in outdoor scenes;

[0063] Step S2: performing illumination distribution statistics on the well-exposed images in the training set to produce a benchmark illumination distribution;

[0064] Step S3: performing a complementary operation on the poorly exposed images in the training set to obtain complementary exposure images;

[0065] Step S4: performing exposure correction on the poorly exposed image and its complementary exposed image based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images;

[0066] Step S5: Design a loss function based on the poorly exposed image and the enhanced corrected image described in step S4, the complementary exposed image and the enhanced corrected image described in step S4, and the reference illumination distribution obtained in step S2. , using the loss function The exposure bias enhancement network model is trained to obtain a final exposure bias enhancement network model after training.

[0067] In this embodiment, the well-exposed images in the training set described in step S1 are used to create the reference illumination distribution, and well-exposed images in various outdoor scenes are preferably selected.

[0068] In this embodiment, the well-exposed image in step S2 is decomposed based on the Retinex operation to obtain its illumination components, perform illumination distribution statistics, and generate a reference illumination distribution. The specific method is as follows:

[0069] The Retinex decomposition operation is inspired by the biological vision model, and the equation is:

[0070] ;

[0071] in, represents RGB channels, It is The illumination component of a well-exposed image, It is Well-exposed images.

[0072] The illumination distribution statistics of the illumination components with good exposure are performed to produce the reference illumination distribution The equation is:

[0073] ;

[0074] Where, The final exposure is good, which will guide the brightness distribution of the final corrected image to approach the good distribution. is the number of well-exposed images.

[0075] In this embodiment, the poorly exposed image in step S3 is subjected to a complementary operation to obtain a complementary exposure image, that is, the poorly exposed image in the training set is subjected to a complementary operation to obtain a complementary exposure image. ,image Based on normalization, the values are Within the scope, through Complementary operation to obtain complementary exposure images ;

[0076] In this embodiment, the loss function described in step S5 Training the exposure bias enhancement network model involves training the entire network model using a backpropagation algorithm and updating its parameters. A determination mechanism is implemented to determine whether a predetermined number of training iterations have been completed. Upon completion of all training iterations, a trained and optimized exposure bias enhancement network model is obtained.

[0077] In this embodiment, after model training is completed, a test set is preferably established to validate the exposure bias enhancement network model of this application. Specifically, a test set is established, consisting of poorly exposed images from outdoor scenes. The poorly exposed images in the test set are input into the exposure bias enhancement network model trained in step S5. The exposure bias enhancement network model processes the input poorly exposed images and outputs a final enhanced image.

[0078] This embodiment of the method is a high-quality exposure deviation enhancement network model construction technology that is applicable to various complex exposures. It effectively alleviates the problem of existing methods' strong dependence on high-quality reference images, overcomes the image degradation problem caused by poor exposure, and produces a benchmark brightness distribution by statistically analyzing a large number of well-exposed images of different scenes to guide high-quality exposure correction in an unsupervised manner.

[0079] like Figure 3 and Figure 4 As shown in FIG. 1 , it is a comparison diagram of an example of correction of the unsupervised exposure method described in this embodiment, that is, Figure 3 middle, Figure 3 (a) is a poorly exposed image, which is unclear due to insufficient exposure. Figure 3 (b) is the image after exposure correction using the exposure deviation enhancement network model described in this embodiment. Figure 3 (a) is obviously clearer in comparison.

[0080] Figure 4 middle, Figure 4 (a) is a poorly exposed image, which is not clear due to overexposure. Figure 4 (b) is the image after exposure correction using the exposure deviation enhancement network model described in this embodiment. Figure 4 (a) is obviously clearer in comparison.

[0081] Implementation 2: This implementation further limits the method for constructing the exposure deviation enhancement network model described in Implementation 1. In this implementation, 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:

[0082] First, the illumination component of the poorly exposed image or its complementary exposed image is obtained based on the Retinex decomposition operation, and then the initial exposure features are obtained through Conv. The initial exposure features are subjected to the illumination-aware three-branch enhancement module to obtain the corrected exposure features. After repeating the correction four times, the poorly exposed image or its complementary exposed image is subjected to channel-attention adaptive compression based on the corrected exposure features to obtain the corresponding enhanced corrected image.

[0083] In this embodiment, if Figure 1 As shown, poorly exposed image Obtain its illumination component based on Retinex decomposition operation and its reflected component , and then get the initial exposure features through Conv , initial exposure characteristics The three-branch enhancement module based on illumination perception obtains the corrected exposure characteristics After repeating the correction 4 times, the exposure features obtained by each correction are spliced together, and the channels are adaptively compressed based on channel attention to obtain the final exposure-enhanced image. .

[0084] In this embodiment, if Figure 1 As shown, the complementary exposure image Obtain its illumination component based on Retinex decomposition operation , and then get the initial exposure features through Conv , initial exposure characteristics The three-branch enhancement module based on illumination perception obtains the corrected exposure characteristics After repeating the correction 4 times, the exposure features obtained by each correction are spliced together, and the channels are adaptively compressed based on channel attention to obtain the final exposure-enhanced image. .

[0085] This embodiment further limits step S4 and illustrates the scheme of step S4. The method corrects poorly exposed images based on the exposure deviation enhancement network model. The three-branch feature extraction module effectively obtains different exposure features. The illumination perception feature fusion module performs fine-grained fusion according to the exposure conditions of the original image. The channel attention is used to adaptively compress the feature channel to obtain a high-quality enhanced image.

[0086] Implementation method 3: This implementation method further limits the exposure deviation enhancement network model construction method described in implementation method 2. In this implementation method, the method of performing the Retinex decomposition operation on the poorly exposed image or its complementary exposed image is the same, wherein the illumination component is obtained by performing the Retinex decomposition on the poorly exposed image. and the reflected component :

[0087]

[0088] ;

[0089] in, represents RGB channels, is a poorly exposed image. .

[0090] This embodiment further limits the exposure bias enhancement network model construction method described in the second embodiment, and explains the Retinex decomposition operation. The 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 illumination component related to the ambient light and a reflectance component related to the inherent characteristics of the image. Improving image quality by adjusting these two components is inspired by the biological vision model.

[0091] Implementation 4: This implementation further limits the exposure deviation enhancement network model construction method described in Implementation 2. In this implementation, the method for obtaining the corrected exposure characteristics by the initial exposure characteristics through the illumination-aware three-branch enhancement module is as follows:

[0092] First, the initial exposure features are extracted based on three-branch feature extraction to obtain input features. , balanced characteristics and complementary features , and then perform feature fusion based on the three illumination perception functions and Conv to finally obtain the corrected exposure features.

[0093] This embodiment further limits the illuminance-aware three-branch enhancement module and explains the illuminance-aware three-branch enhancement module, which includes two parts: three-branch feature extraction and illuminance-aware feature fusion. This method obtains the corrected exposure features, and prepares for the subsequent use of channel attention to adaptively compress feature channels to obtain high-quality enhanced images.

[0094] Implementation 5: This implementation further limits the exposure bias enhancement network model construction method described in Implementation 4. In this implementation, the three-branch feature extraction is as follows:

[0095] First, the initial exposure feature is learned based on the shared weight dual-branch convolution designed in parallel by standard Conv and center difference convolution CDConv. and its complementary characteristics At the same time, the exposure feature estimates its balanced feature based on instance normalization IN .

[0096] In this embodiment, if Figure 2 As shown, the exposure characteristics and its complementary exposure characteristics Based on the shared weight dual-branch convolution activation and learning with opposite exposure features, the input features are obtained and complementary features ,at the same time Based on instance normalization (IN), the inherent features of the image such as structure and texture are better preserved to obtain balanced features. , the input features obtained by three-branch feature extraction , complementary features and balanced characteristics exist Under guidance, feature fusion is performed based on three illumination perception fusion functions and Conv, and finally the corrected exposure features are obtained.

[0097] The purpose of the dual-branch convolution is to enhance the model's perception of image details and textures, thereby extracting local detail perception of different exposure features and improving the discriminability of exposure feature expression. The dual-branch convolution consists of a standard Conv block and a center difference convolution block (CDConv) in parallel. It uses the features output by channel splicing, and the 1x1 Conv fusion feature is compressed to restore the number of channels. The center difference convolution enhances the ability to capture local structures such as edges and corners, is more robust to noise, and can suppress the influence of illumination changes or uniform noise.

[0098] This embodiment further limits the method for constructing an exposure bias enhancement network model and illustrates three-branch feature extraction, which effectively obtains different exposure features.

[0099] Implementation 6. In this implementation, this implementation further limits the exposure deviation enhancement network model construction method described in Implementation 4. In this implementation, the feature fusion based on the three illumination perception functions and Conv is as follows:

[0100] First, the three illumination perception function formulas are:

[0101] ;

[0102] ;

[0103] ;

[0104] in, is the input feature extracted by the three-branch feature, It is the complementary feature extracted by the three-branch feature. It is the balanced feature extracted from the three-branch feature. is the illumination component of the input image, and the input features and the equalization features are based on Function and Conv are fused to obtain fusion features , complementary features and balanced features are based on Function and Conv are fused to obtain fusion features , fusion features and based on Function and Conv fusion to obtain enhanced features .

[0105] This embodiment further limits the exposure deviation enhancement network model construction method described in the fourth embodiment, and describes the feature fusion of three illumination perception functions and Conv. The enhanced features obtained by this method are: It improves the loss of color and structure caused by overexposure, reduces the noise introduced by underexposure, and effectively reveals the color and texture details in dim areas.

[0106] Implementation 7: This implementation further limits the exposure bias enhancement network model construction method described in Implementation 1. In this implementation, in step S5, the loss function is:

[0107] ;

[0108] in, is the latent factor consistency loss, is the brightness distribution loss, It is the structural similarity loss.

[0109] In this embodiment, the loss function specifically adopts two widely recognized factor decomposition standards, Retinex theory and Fourier transform, to formulate the latent factor consistency loss ;

[0110] Retinex theory believes that the reflection component of the image should remain unchanged during the enhancement process. and its enhanced image EI share the same reflection component, yes The enhanced image, and With consistent reflection components, 、 as well as Obtain the corresponding reflection component based on the Retinex decomposition operation 、 as well as , the loss is calculated by the MSE loss function (mean square error loss) , ensuring the consistency of color tone, structure and texture of the enhanced image, is defined as:

[0111] ;

[0112] The amplitude and phase components of the images with different exposures are obtained by Fourier transform, which show similar Fourier frequency characteristics. Specifically, the amplitude component mainly contains brightness information, while the phase component is more closely related to structural details. and its enhanced image EI should have similar phase components, yes The enhanced image, and have similar phase components, 、 as well as Calculate the corresponding phase component based on Fourier transform 、 as well as ,pass The loss function (absolute value loss function) calculates the loss , further ensures the consistency of enhanced image detail texture and structural contour, which is defined as:

[0113] ;

[0114] Therefore, the above loss function can be summarized as:

[0115] ;

[0116] The well-exposed image has similar illumination distribution to guide the brightness distribution of the enhanced image, and the enhanced image is obtained based on the Retinex decomposition operation. Lighting component And calculate its brightness distribution , design brightness distribution loss using KL divergence and Wasserstein distance (EMD) , KL divergence makes the two distributions overlap as much as possible, ensuring that the brightness distribution of the enhanced image is similar to the baseline distribution , combined with EMD, can promote more adaptive exposure enhancement based on the intrinsic characteristics of the image, which is defined as:

[0117] ;

[0118] Therefore, the above loss function can be summarized as:

[0119] ;

[0120] In the formula is the total loss function, is the structural similarity loss.

[0121] This embodiment further limits the exposure bias enhancement network model construction method and explains the loss function. The latent factor consistency loss , ensure 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 deviation and enhancing the generalization ability of the network. Brightness distribution loss , ensuring that the enhanced image presents high quality and balanced brightness distribution. It is a structural similarity loss, which ensures the structural quality of the enhanced image. In summary, this loss function can be better used for the exposure bias enhancement network model training of this application.

Claims

1. A method for constructing an exposure bias enhancement network model for unsupervised exposure correction, characterized in that: The following steps are involved: Step S1: obtaining a training set, where the training set includes poorly exposed images and well-exposed images in outdoor scenes; Step S2: performing illumination distribution statistics on the well-exposed images in the training set to produce a benchmark illumination distribution; Step S3: performing a complementary operation on the poorly exposed images in the training set to obtain complementary exposure images; Step S4: performing exposure correction on the poorly exposed image and its complementary exposed image based on the exposure deviation enhancement network model to obtain corresponding enhanced corrected images; specifically, First, the illumination component of the poorly exposed image or its complementary exposed image is obtained based on the Retinex decomposition operation. Then, the initial exposure features are obtained through Conv. The initial exposure features are then subjected to the illumination-aware three-branch enhancement module to obtain the corrected exposure features (Concat). After repeating the correction four times, the iteratively corrected exposure features (Concat) are obtained. Then, based on the iteratively corrected exposure feature Concat, a corresponding enhanced corrected image is obtained; The process of obtaining the corrected exposure characteristics from the initial exposure characteristics through the illumination perception three-branch enhancement module is as follows: First, the initial exposure features are extracted based on three-branch features to obtain input features. , balanced characteristics and complementary features , then perform feature fusion based on the three illumination perception functions and Conv, and finally obtain the corrected exposure feature Concat; Step S5: Design a loss function based on the poorly exposed image and the enhanced corrected image described in step S4, the complementary exposed image and the enhanced corrected image described in step S4, and the reference illumination distribution obtained in step S2. , using the loss function The exposure bias enhancement network model is trained to obtain a final exposure bias enhancement network model after training.

2. The exposure bias enhancement network model construction method according to claim 1, characterized in that: The process of obtaining the corresponding enhanced corrected image based on the iteratively corrected exposure feature Concat in step S4 is as follows: Based on the iteratively corrected exposure feature Concat, channel attention adaptive compression is performed on the poorly exposed image or its complementary exposed image to obtain a corresponding enhanced corrected image.

3. The exposure bias enhancement network model construction method according to claim 1, characterized in that: The method of performing the Retinex decomposition operation on the poorly exposed image or its complementary exposed image is the same, wherein the illumination component is obtained by decomposing the poorly exposed image based on Retinex. and the reflected component : ， in, represents RGB channels, is a poorly exposed image. .

4. The exposure bias enhancement network model construction method according to claim 1, characterized in that: The three-branch feature extraction method is: First, the initial exposure feature is learned based on the shared weight dual-branch convolution designed in parallel by standard Conv and center difference convolution CDConv. and its complementary characteristics At the same time, the exposure feature estimates its balanced feature based on instance normalization IN .

5. The exposure bias enhancement network model construction method according to claim 1, characterized in that: The method for feature fusion based on three illumination perception functions and Conv is: First, the three illumination perception functions are: ； ； ； in, is the illumination component of the input image, and the input feature and balanced characteristics based on Function and Conv are fused to obtain fusion features , complementary features and balanced characteristics based on Function and Conv are fused to obtain fusion features , fusion features and based on Function and Conv fusion to obtain enhanced features .

6. The exposure bias enhancement network model construction method according to claim 1, characterized in that: In the step S5, the loss function for: ； in, is the latent factor consistency loss, is the brightness distribution loss, It is the structural similarity loss.

7. An unsupervised exposure correction system, characterized in that Includes the following modules: Module 1: used to obtain a training set, which includes poorly exposed images and well-exposed images of outdoor scenes; Module 2: Used to calculate the illumination distribution statistics of well-exposed images in the training set and create a benchmark illumination distribution; Module 3: Used to perform complementary operations on poorly exposed images in the training set to obtain complementary exposure images; Module 4: for performing exposure correction on the poorly exposed image and its complementary exposed image based on an exposure deviation enhancement network model to obtain corresponding enhanced corrected images; The module 4 further includes: Submodule 1: Used to obtain the illumination component of the poorly exposed image or its complementary exposed image based on the Retinex decomposition operation, then obtain the initial exposure feature through Conv, and obtain the corrected exposure feature Concat after the initial exposure feature is processed through the illumination-aware three-branch enhancement module. After repeating the correction four times, the iteratively corrected exposure feature Concat is obtained. Submodule 2: for obtaining a corresponding enhanced corrected image based on the iteratively corrected exposure feature Concat; The submodule 1 further includes a unit for extracting input features based on the three-branch feature extraction of the initial exposure features. , balanced characteristics and complementary features , then perform feature fusion based on the three illumination perception functions and Conv, and finally obtain the corrected exposure feature Concat; Module 5: Designing a loss function based on the poorly exposed image and its enhanced corrected image described in Module 4, the complementary exposed image and its enhanced corrected image described in Module 5, and the reference illumination distribution obtained in Module 2 , using the loss function The exposure bias enhancement network model is trained to obtain a final exposure bias enhancement network model after training.

8. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the steps of the method according to any one of claims 1 to 6 when executing a program stored in a memory.

9. 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 a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images