An unsupervised exposure correction method

Abstract

The application discloses a kind of unsupervised exposure correction methods, method includes: constructing semantic prompt guide's exposure perception module, the exposure level of image is perceived at visual semantic level;Exposure parameter modulation is constructed pseudo-supervision generation module, according to image exposure level adaptive exposure modulation is carried out, and synthesis is used to guide model training pseudo-supervision image;Construct color degradation compensation exposure correction module, by joint execution luminance correction and color compensation, the luminance distortion in abnormal exposure image and color degradation are recovered;Unsupervised exposure correction network is trained using the network optimization mechanism of multiple loss functions;Based on the unsupervised exposure correction network after training, input image is handled, and image luminance level and color fidelity are effectively improved.The application generates pseudo-supervision image according to image exposure level adaptive to guide model training, and compensates the color information of image while luminance correction, to realize the collaborative recovery of luminance and color.

Description

Technical Field

[0001] This invention relates to the fields of image enhancement and computer vision, and in particular to an unsupervised exposure correction method. Background Technology

[0002] With the development of digital imaging technology and smart terminal devices, the scale of image data continues to grow, making image acquisition and dissemination more convenient. However, due to variations in lighting and limitations of imaging equipment, images captured by sensors are prone to overexposure or underexposure, leading to a decline in visual quality. Exposure correction technology aims to effectively correct abnormally exposed images to restore natural brightness distribution and true color reproduction, and has been widely used in fields such as smart security and intelligent manufacturing. Conducting research on exposure correction technology has significant research importance and application value.

[0003] Early exposure correction methods primarily relied on manually designed exposure adjustment strategies, such as histogram equalization and tone mapping. While simple to implement, these methods, due to their reliance on fixed rules, struggled to handle complex and spatially uneven exposure degradation. With the development of deep learning, supervised learning-based exposure correction methods have significantly improved correction performance by learning the mapping relationship between abnormally exposed and normally exposed images. However, these methods depend on large amounts of high-quality paired training data, which is costly and difficult to acquire, limiting the practical application of supervised methods. Therefore, how to achieve accurate brightness and color correction without labeled data requires further exploration.

[0004] In recent years, researchers have made initial explorations into unsupervised exposure correction tasks. These methods typically construct pseudo-supervisory signals based on statistical priors or pre-defined transformation functions to guide the network in learning exposure correction mapping relationships, thereby achieving effective correction of abnormally exposed images without the constraints of pairwise data. For example, Nguyen et al. constructed pseudo-supervised images through brightness mapping and multi-exposure fusion, achieving more robust brightness recovery under extreme exposure conditions. Cui et al., based on the principle of radiometric imaging, modeled exposure correction as pixel-level exposure transformation, thus preserving image detail structure. However, existing methods often rely on predefined correction rules when constructing pseudo-supervisory signals, making it difficult to adapt to the complex and varied exposure degradation in input images, thus affecting the correction effect. Furthermore, existing methods often focus on restoring image brightness, neglecting the color degradation caused by abnormal exposure, resulting in insufficient color saturation in the corrected image and affecting visual quality. Summary of the Invention

[0005] This invention provides an unsupervised exposure correction method. The method adaptively generates pseudo-supervised images based on the image's exposure level to guide model training, and compensates for the image's color information while correcting brightness, thereby achieving coordinated restoration of brightness and color. Details are described below:

[0006] An unsupervised exposure correction method, the method comprising:

[0007] An unsupervised exposure correction network is constructed, consisting of an exposure perception module guided by semantic prompts, a pseudo-supervised generation module for exposure parameter modulation, and an exposure correction module for color degradation compensation.

[0008] An unsupervised exposure correction network is trained using a network optimization mechanism with multiple loss functions.

[0009] The input image is processed using a trained unsupervised exposure correction network to achieve coordinated restoration of brightness and color.

[0010] The semantically prompted exposure perception module establishes a relationship between image content and exposure status at the visual semantic level, and generates an exposure level representation of the image by combining brightness statistics.

[0011] Construct exposure text cue pairs for underexposure and overexposure respectively, to describe the exposure characteristics of an image that is generally dark with deep shadows and generally bright with prominent highlights;

[0012] The text prompts are input into the pre-trained cross-modal model's text encoder to obtain the corresponding text feature representations; the input image is fed into the cross-modal model's image encoder to extract the global feature representations of the image, and similarity calculations are performed with the underexposed and overexposed text features respectively to obtain the underexposed semantic score and overexposed semantic score of the image at the global level.

[0013] The input image is divided into image patches of corresponding scales at multiple scales, and the similarity between each image patch and the text prompt is calculated to obtain the corresponding local level underexposure semantic score and overexposure semantic score.

[0014] The semantic scores at the global and local scales are uniformly mapped to the finest spatial resolution and then stitched together to construct a multi-scale exposure score vector. The mean and variance of the brightness of each image patch are calculated and stitched together with the multi-scale exposure score vector to obtain the exposure level representation of each image patch.

[0015] The pseudo-supervision generation module for exposure parameter modulation is as follows:

[0016] The exposure level representation of each image patch is added to the type embedding, and after linear mapping, it is added to the position encoding to obtain the feature representation. The feature representation corresponding to each image patch is input into the Transformer network to predict the exposure adjustment parameters.

[0017] From the intervals respectively and Random sampling A disturbance value, denoted as The perturbation value is applied to the exposure adjustment parameters to generate a set of candidate adjustment parameters;

[0018] Based on the exposure adjustment parameters and candidate adjustment parameters, gamma correction is used to modulate the exposure of the input image, generating multiple images to form a reference image set. Each image in the reference image set is divided into non-overlapping pixel local regions, and the saturation, contrast, and gray-level distribution skewness of each local region are calculated. For the same spatial location, the local region with the highest quality score is selected from the reference image set as the pseudo-supervision content for that location. After traversing all spatial locations, the pseudo-supervision image is obtained.

[0019] The exposure correction module for color degradation compensation is as follows:

[0020] Define low-intensity boundary response function High-intensity boundary response function and pixels The passageway is extremely poor. ;

[0021] Constructing a color deviation indicator based on boundary response and channel range At this time, the two boundary responses are combined in a mutually inhibiting manner:

[0022] ;

[0023] Image-level degradation coefficients Pixel-level color deviation indicator Combined, a color compensation map is obtained. The formula is expressed as:

[0024]

[0025] in, Indicates the color compensation map at the pixel level. The value at that location, This indicates the softplus activation function; it adjusts the brightness of the image. Convert to Lab color space to obtain the luminance channel L and color channels a and b, and then use the color compensation map. The formula for compensating color channels is as follows:

[0026] .

[0027] The network optimization mechanism employing multiple loss functions to train the unsupervised exposure correction network is as follows:

[0028] We design a network optimization mechanism based on multiple loss functions, with loss functions designed for the pseudo-supervised image construction process and the exposure correction process, respectively, to achieve joint optimization. Specifically, we design reconstruction loss and color deviation loss for the pseudo-supervised image construction process, and brightness correction loss and saturation perception loss for the exposure correction process.

[0029] Wherein, the color deviation loss is:

[0030] For the finest scale, the first For each image patch, the color deviation is defined as:

[0031]

[0032] in, , and They represent the first The average value of an image patch across the R, G, and B channels;

[0033] Color deviation loss Defined as:

[0034]

[0035] in, and These represent the input images respectively. Compared with reference image At the finest scale Image blocks, For image blocks Color deviation, For image blocks Color deviation.

[0036] The method further includes defining the channel overflow amplitude of the output image as:

[0037]

[0038] in, , and They represent medium pixel The normalized values ​​on the three RGB channels define the color amplitude in the Lab color space. Saturation perception loss Represented as:

[0039] ;

[0040] in, Represents the image region. Represents color compensation image The mean, These are the weight parameters.

[0041] The beneficial effects of the technical solution provided by this invention are:

[0042] 1. This invention utilizes cross-modal semantic information to estimate the exposure level of an image and adaptively constructs a pseudo-supervised image to improve the reliability of the guidance information; at the same time, by quantifying the degree of color degradation of the image and applying color compensation, it effectively restores the true color of the image while restoring the image brightness.

[0043] 2. The proposed solution can improve the stability and authenticity of exposure correction results without the need for paired training data, and can then be applied to practical scenarios such as intelligent monitoring and autonomous driving.

[0044] 3. This invention can effectively correct underexposed and overexposed images without the need for paired training data, thereby reducing the reliance on large-scale manually labeled data and improving the flexibility and practicality of exposure correction.

[0045] 4. This invention constructs an exposure perception module and a pseudo-supervised image generation module based on cross-modal semantic alignment. By perceiving the exposure level of the input image, a higher quality pseudo-supervised image is constructed for constrained training.

[0046] 5. This invention constructs a color compensation and exposure correction module, which improves the visual quality of the correction results by jointly restoring image brightness and compensating for color degradation caused by abnormal exposure.

[0047] 6. This invention designs a network optimization mechanism based on multiple loss functions, which effectively constrains the pseudo-supervised construction process and the exposure correction process, thereby improving the training effect and final performance of the entire unsupervised exposure correction network. Attached Figure Description

[0048] Figure 1 A flowchart of an unsupervised exposure correction method;

[0049] Figure 2 This is a visualization of the method proposed in this invention on the MSEC dataset. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0051] I. Constructing an Exposure Awareness Module Guided by Semantic Prompts

[0052] The exposure states of different input images often vary, and accurately representing the exposure level of an image is crucial for constructing high-quality pseudo-supervised learning. By leveraging the image-text semantic alignment capabilities of cross-modal models, a correspondence between image content and exposure text can be established, thereby enabling the perception of the image's exposure state. Based on this, this embodiment of the invention constructs a semantically prompted exposure perception module. By establishing a correlation between image content and exposure state at the visual semantic level and combining it with brightness statistics, it generates a representation of the image's exposure level.

[0053] Specifically, firstly, corresponding exposure text prompts are constructed for the two typical exposure states of underexposure and overexposure. This is used to describe the exposure characteristics of an image that is generally dark with deep shadows or generally bright with prominent highlights. In embodiments of the present invention, "Generally underexposed image with overall darkshadows", "Generally overexposed image with overall bright highlights". The text prompt is input to the pre-trained cross-modal model's text encoder to obtain the corresponding text feature representation.

[0054] Secondly, the input image The image encoder, fed into a cross-modal model, extracts the global feature representation of the image and calculates similarity with the underexposed and overexposed text features respectively, to obtain the underexposed semantic score of the image at the global level. semantic score of overexposure This process can be expressed by the following formula:

[0055] (1)

[0056] in, and These represent the image encoder and text encoder in the cross-modal model, respectively, used to map the input image and exposure text cue to the same semantic feature space. Represents cosine similarity. This refers to the temperature coefficient. In this embodiment of the invention, the cross-modal model is CLIP, the image encoder uses a ViT-B / 32 architecture, and the text encoder uses a Transformer architecture. Set to 0.02.

[0057] Considering that the exposure levels may differ in different spatial regions of an image, the input image is divided into multiple scales to achieve local exposure perception. Specifically, the following steps are taken: The local scale is then used to input the image at the _ local scale, and the input image is at the _ local scale. Divided into several scales A number of non-overlapping image blocks.

[0058] In this embodiment of the invention, Set to 2, corresponding to and The division method, The similarity score is 16. For each image patch, the similarity between it and the text prompt is calculated to obtain the corresponding local underexposure and overexposure semantic scores, denoted as [missing information]. and ,in Indicates the first Image patch index at each scale.

[0059] Subsequently, the semantic scores at the global and local scales are uniformly mapped to the finest spatial resolution and then stitched together to construct a multi-scale exposure score vector. For the finest scale... The corresponding multi-scale exposure score vector for each image patch is represented as follows:

[0060] (2)

[0061] in, This indicates that the first local scale contains the first... The corresponding block index of the finest-scale image patch. and These are the semantic scores for underexposure and overexposure of the corresponding image patch, respectively. and The first At the local scale, the first Semantic scores for underexposure and overexposure of individual image patches.

[0062] Finally, the average brightness of each image patch is calculated. and variance This is then concatenated with the multi-scale exposure score vector to obtain the exposure level representation for each image patch. .

[0063] II. Constructing a pseudo-supervisory generation module for exposure parameter modulation

[0064] Subsequently, this embodiment of the invention constructs a pseudo-supervised generation module for exposure parameter modulation. By using the exposure level representation to predict the exposure adjustment parameters, the input image is subjected to exposure modulation to adaptively generate a reference image, and then a pseudo-supervised image is synthesized to guide model training.

[0065] Specifically, the exposure level of each image patch is first represented. The feature representation is obtained by incorporating type embedding, applying a linear mapping, and then adding it to the positional encoding. Subsequently, the feature representations corresponding to each image patch are... Input the Transformer network to predict exposure adjustment parameters The formula is expressed as:

[0066] (3)

[0067] (4)

[0068] in, Indicates type embedding, used for distinction The different components of This represents a linear mapping operation. Indicates the first Position encoding of each image patch Indicates the first Number of image patches at a local scale This refers to the Transformer encoder network. In this embodiment of the invention, the Transformer encoder network includes four Transformer encoder layers, each containing four attention heads, and outputs exposure adjustment parameters corresponding to each image patch through a linear prediction layer.

[0069] Next, adjust the parameters based on the predicted exposure. Gamma correction is used for the input image Exposure modulation is performed to generate a reference image. The formula is expressed as:

[0070] (5)

[0071] in, This indicates the exposure adjustment parameters for each image block. The exposure adjustment parameter diagram is obtained by expanding according to the corresponding spatial location.

[0072] To further enhance the diversity of pseudo-supervision, from the intervals and Random sampling A disturbance value, denoted as ,in In this embodiment of the invention, Set it to 0.5, Set to 3. Apply the perturbation value to the exposure adjustment parameter. A set of candidate adjustment parameters is generated, expressed by the formula:

[0073] (6)

[0074] Based on the candidate adjustment parameters, gamma correction is also used on the input image. Exposure modulation is performed to generate multiple candidate images. Subsequently, the reference images and candidate images are used together as a reference image set, and local quality assessment is performed on each image in this set.

[0075] Specifically, each image in the reference image set is divided into non-overlapping segments. Calculate the saturation of each local region of a pixel. Contrast and grayscale distribution skewness The local quality score is defined as:

[0076] (7)

[0077] in, Assign a quality score to the local region. For the same spatial location, select the local region with the highest quality score from the reference image set as the pseudo-supervision content for that location. After traversing all spatial locations, obtain the pseudo-supervision image. .

[0078] III. Constructing an Exposure Correction Module for Color Degradation Compensation

[0079] This invention constructs an exposure correction module for color degradation compensation. By jointly performing brightness correction and color compensation, it restores the brightness distortion and color degradation in abnormally exposed images, thereby generating the final correction result.

[0080] Specifically, first, the input image The image features are input into a feature extraction network to obtain image features. This feature extraction network adopts the lightweight U-Net structure known in the art, removing the output layer and retaining only the encoding, decoding, and skip connection structures. Subsequently, the extracted image features are input into the luminance branch. The luminance branch consists of a luminance output layer, an exponential mapping operation, and a gamma correction operation. The luminance output layer uses the same convolutional mapping structure as the output layer of the lightweight U-Net to generate the original luminance response map. The exponential mapping operation is used to transform the original luminance response map. Convert to brightness correction parameter diagram The formula is expressed as:

[0081] (8)

[0082] in, For pixel position, and These represent the original luminance response map and the luminance correction parameter map at the pixel level, respectively. The value at that location, This represents an exponential function. Finally, the brightness correction parameter diagram is used. For the input image Perform gamma correction to obtain a brightness-corrected image. The formula is expressed as:

[0083] (9)

[0084] Because underexposure or overexposure causes pixel values ​​to shift towards the boundaries of the dynamic range, some RGB channels may approach or reach their value boundaries, resulting in channel clipping. This disrupts the original proportional relationship between channels, causing color degradation. To characterize the relative degree of color degradation at different locations in an image, this invention employs a pixel-level color deviation indicator. In an abnormally exposed input image, the closer the channel value is to the value boundary, the more likely the pixel has been affected by channel clipping and exhibits more obvious color degradation.

[0085] Therefore, in order to measure the relationship between RGB channel values ​​and their value boundaries, low-intensity boundary response functions are defined respectively. and high-intensity boundary response function The formula is expressed as:

[0086] (10)

[0087] in, , and Representing pixels Normalized values ​​across the three RGB channels. The larger the value, the closer the pixel is to the lower limit of the value in at least one channel, and the higher the possibility of underexposure causing color degradation; The larger the value, the closer the pixel is to the upper limit of the value in at least one channel, and the higher the possibility of color degradation caused by overexposure.

[0088] Furthermore, considering the imbalance between channels, pixels are defined. The channel range at that location is:

[0089] (11)

[0090] in, It is used to describe the degree of difference between the RGB channels within the same pixel, reflecting the uneven effect of abnormal exposure on different channels. In addition, when the RGB channels of a pixel are close to different value boundaries, the pixel corresponds to an inherently high saturation region rather than a color degradation region, but it still exhibits a high boundary response.

[0091] Therefore, to exclude the above situation, a color deviation indicator is constructed based on the boundary response and channel range. When the two boundary responses are combined in a mutually inhibiting manner, the formula is expressed as:

[0092] (12)

[0093] Furthermore, image features extracted by the feature extraction network are input into the color branch. The color branch consists of a convolutional mapping layer and a global average pooling operation, where the convolutional mapping layer generates a single-channel color response, and the global average pooling operation converts this response into image-level degradation coefficients. This adjusts the overall compensation intensity of the image. The image-level degradation coefficient is then adjusted. Pixel-level color deviation indicator Combined, a color compensation map is obtained. The formula is expressed as:

[0094] (13)

[0095] in, Indicates the color compensation map at the pixel level. The value at that location, This indicates the softplus activation function.

[0096] Finally, to avoid color compensation affecting the brightness correction results, the brightness-corrected image is... Convert to Lab color space to obtain the luminance channel L and color channels a and b, and then use the color compensation map. The formula for compensating color channels is as follows:

[0097] (14)

[0098] By compensating the color channels , With brightness channel Recombined and converted back to linear RGB space, the final output image is obtained. The formula is expressed as:

[0099] (15)

[0100] in, This indicates a conversion operation from the Lab color space to the linear RGB color space.

[0101] IV. Designing a network optimization mechanism based on multiple loss functions

[0102] This invention proposes a network optimization mechanism based on multiple loss functions, which designs loss functions for the pseudo-supervised image construction process and the exposure correction process respectively, in order to achieve joint optimization.

[0103] First, a reconstruction loss is designed for the pseudo-supervised image construction process. and color deviation loss .

[0104] Reconstruction losses Defined as a reference image With pseudo-supervised images The mean square error between them is expressed by the formula:

[0105] (16)

[0106] in, Represents the image region. Indicates the pixel position in the image. express The square of the norm.

[0107] For the finest scale, the first For each image patch, the color deviation is defined as:

[0108] (17)

[0109] in, , and They represent the first The average value of an image patch across the R, G, and B channels.

[0110] Color deviation loss Defined as:

[0111] (18)

[0112] in, and These represent the input images respectively. Compared with reference image At the finest scale Image blocks, For image blocks Color deviation, For image blocks Color deviation.

[0113] Secondly, a brightness correction loss is designed for the exposure correction process. and saturation perception loss .

[0114] Among them, brightness correction loss It consists of a mean squared error term and a total variational constraint term, and the formula is expressed as:

[0115] (19)

[0116] in, and Let represent the gradient operators in the horizontal and vertical directions, respectively. and These represent the brightness-corrected image and the pseudo-supervised image at the pixel level, respectively. The value at that location, The brightness correction parameter map representing the brightness branch prediction at the pixel level The value at that location, These are the weight parameters for the total variational constraint term. In this embodiment of the invention, Set it to 500.

[0117] In addition, output image The channel overflow amplitude is defined as:

[0118] (20)

[0119] in, , and They represent medium pixel Normalized values ​​across the three RGB channels. Simultaneously, color amplitude is defined in the Lab color space. Saturation perception loss Represented as:

[0120] (twenty one)

[0121] in, Represents color compensation image The mean, This is a weighting parameter. In this embodiment of the invention, Set to 0.01.

[0122] Finally, summing up the above losses forms the overall optimization objective function:

[0123] (twenty two)

[0124] In this embodiment of the invention, , , All are set to 1.

[0125] V. Training an Unsupervised Exposure Image Correction Network

[0126] The network proposed in this embodiment of the invention employs an end-to-end unsupervised training approach. The network includes a semantically prompted exposure-aware module, a pseudo-supervised generation module for exposure parameter modulation, and an exposure correction module for color degradation compensation. During the training phase, the pre-trained cross-modal model parameters in the exposure-aware module remain frozen and do not participate in network parameter updates; the remaining learnable modules are jointly trained through a multi-loss function optimization mechanism, and employ... The network is optimized using this loss function until convergence.

[0127] During the inference phase, the input abnormally exposed image directly obtains the correction result through the exposure correction module with color degradation compensation. The method proposed in this invention achieves effective brightness and color correction for images at different exposure levels through an exposure-adaptive pseudo-supervised image generation strategy and a targeted color compensation mechanism. This method can be effectively deployed in typical application scenarios such as smart security, intelligent manufacturing, and autonomous driving, thereby improving the visual perception quality under complex lighting conditions and further enhancing the overall accuracy of related downstream visual tasks.

[0128] VI. Implementation Results

[0129] To verify the effectiveness of the embodiments of the present invention, a comparative experiment was conducted on the exposure correction dataset MSEC, and the visualization results are as follows. Figure 2 As shown. Compared with the method of Cui et al., the embodiments of the present invention can more effectively improve the brightness level and color fidelity of images when correcting underexposed or overexposed images, thereby obtaining higher quality corrected images.

[0130] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An unsupervised exposure correction method, characterized in that, The method includes: An unsupervised exposure correction network is constructed, consisting of an exposure perception module guided by semantic prompts, a pseudo-supervised generation module for exposure parameter modulation, and an exposure correction module for color degradation compensation. An unsupervised exposure correction network is trained using a network optimization mechanism with multiple loss functions. The input image is processed using a trained unsupervised exposure correction network to achieve coordinated restoration of brightness and color.

2. The unsupervised exposure correction method according to claim 1, characterized in that, The semantically prompted exposure perception module establishes a relationship between image content and exposure status at the visual semantic level, and generates an exposure level representation of the image by combining brightness statistics. Construct exposure text cue pairs for underexposure and overexposure respectively, to describe the exposure characteristics of an image that is generally dark with deep shadows and generally bright with prominent highlights; The text prompts are input into the pre-trained cross-modal model's text encoder to obtain the corresponding text feature representations; the input image is fed into the cross-modal model's image encoder to extract the global feature representations of the image, and similarity calculations are performed with the underexposed and overexposed text features respectively to obtain the underexposed semantic score and overexposed semantic score of the image at the global level. The input image is divided into image patches of corresponding scales at multiple scales, and the similarity between each image patch and the text prompt is calculated to obtain the corresponding local level underexposure semantic score and overexposure semantic score. The semantic scores at the global and local scales are uniformly mapped to the finest spatial resolution and then stitched together to construct a multi-scale exposure score vector. The mean and variance of brightness are calculated for each image patch, and then concatenated with the multi-scale exposure score vector to obtain the exposure level representation of each image patch.

3. The unsupervised exposure correction method according to claim 1, characterized in that, The pseudo-supervision generation module for the exposure parameter modulation is: The exposure level representation of each image patch is added to the type embedding, and after linear mapping, it is added to the position encoding to obtain the feature representation. The feature representation corresponding to each image patch is input into the Transformer network to predict the exposure adjustment parameters. From the intervals respectively and Random sampling A disturbance value, denoted as The perturbation value is applied to the exposure adjustment parameters to generate a set of candidate adjustment parameters; Based on the exposure adjustment parameters and candidate adjustment parameters, gamma correction is used to modulate the exposure of the input image, generating multiple images to form a reference image set. Each image in the reference image set is divided into non-overlapping pixel local regions, and the saturation, contrast, and gray-level distribution skewness of each local region are calculated. For the same spatial location, the local region with the highest quality score is selected from the reference image set as the pseudo-supervision content for that location. After traversing all spatial locations, the pseudo-supervision image is obtained.

4. The unsupervised exposure correction method according to claim 1, characterized in that, The exposure correction module for color degradation compensation is: Define low-intensity boundary response function High-intensity boundary response function and pixels The passageway is extremely poor. ; Constructing a color deviation indicator based on boundary response and channel range At this time, the two boundary responses are combined in a mutually inhibiting manner: ； Image-level degradation coefficients Pixel-level color deviation indicator Combined, a color compensation map is obtained. The formula is expressed as: ； in, Indicates the color compensation map at the pixel level. The value at that location, This indicates the softplus activation function; it adjusts the brightness of the image. Convert to Lab color space to obtain the luminance channel L and color channels a and b, and then use the color compensation map. The formula for compensating color channels is as follows: 。 5. The unsupervised exposure correction method according to claim 1, characterized in that, The network optimization mechanism employing multiple loss functions for training the unsupervised exposure correction network is as follows: We design a network optimization mechanism based on multiple loss functions, with loss functions designed for the pseudo-supervised image construction process and the exposure correction process, respectively, to achieve joint optimization. Specifically, we design reconstruction loss and color deviation loss for the pseudo-supervised image construction process, and brightness correction loss and saturation perception loss for the exposure correction process.

6. The unsupervised exposure correction method according to claim 5, characterized in that, The color deviation loss is: For the finest scale, the first For each image patch, the color deviation is defined as: ； in, , and They represent the first The average value of an image patch across the R, G, and B channels; Color deviation loss Defined as: ； in, and These represent the input images respectively. Compared with reference image At the finest scale Image blocks, For image blocks Color deviation, For image blocks Color deviation.

7. The unsupervised exposure correction method according to claim 5, characterized in that, The method further includes: defining the channel overflow amplitude of the output image as: ； in, , and They represent medium pixel The normalized values ​​on the three RGB channels define the color amplitude in the Lab color space. Saturation perception loss Represented as: ； in, Represents the image region. Represents color compensation image The mean, These are the weight parameters.

Images