An image exposure correction method based on deep curve estimation

Abstract

The application discloses an image exposure correction method based on deep curve estimation. First, a pair of image exposure correction data sets are constructed, images are shot by using a camera to determine optimal exposure settings, and an image post-processing software is applied to render optimal exposure images in RAW format at different exposures to obtain overexposed and underexposed images. Then, based on the constructed data sets, a deep curve estimation network is trained to determine the parameter settings of the image-specific exposure correction curve. Finally, the parameters estimated by the deep curve estimation network are substituted into a composite exposure correction model composed of a gamma transformation and a logistic function, and the model is used to correct the image exposure. The image exposure correction method proposed in the application realizes accurate general image exposure correction. Compared with other methods on an internationally common data set, the result shows that the method can obtain a more superior exposure correction effect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of digital image processing, computer vision and image enhancement, and particularly relates to an image exposure correction method based on deep curve estimation. BACKGROUND

[0002] When taking a photo, the camera needs to optimize the exposure parameters to ensure the image quality. However, due to the influence of complex lighting conditions, human factors and other factors, the camera often takes photos with suboptimal exposure settings, resulting in overexposure or underexposure problems of the image. Overexposure will cause part of the image pixels to be saturated, resulting in loss of image bright part details and excessive brightness; underexposure will cause part of the image pixels to be insufficiently responsive, resulting in overall image darkness and loss of dark part details. The above exposure problems limit the further application of the image, and exposure correction is an effective technical means to solve the image exposure problem.

[0003] Currently, researchers in the field of exposure correction mainly focus on the correction problem of underexposed images and have obtained rich research results, while there are few existing results for general methods that can simultaneously process overexposed and underexposed images, and the accuracy of existing methods needs to be further improved. In addition, the existing general methods mainly use deep learning technology to construct a large-scale encoder-decoder structure for exposure correction, and such models completely rely on data fitting for modeling, and the method has poor interpretability. For the defects of the above existing image exposure correction methods, the academic and industrial circles at home and abroad have not proposed a good response method. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application proposes an image exposure correction method based on deep curve estimation. This method first constructs an image exposure correction dataset using a camera and image post-processing software, optimizes the exposure settings of the camera by artificial shooting to obtain an optimal exposure image, and renders the RAW format optimal exposure image at different exposure levels to obtain overexposed and underexposed images; then, based on the created dataset, a convolutional neural network is constructed to estimate the parameters of the exposure correction curve from the image and determine the unique exposure correction curve of the image; finally, the curve estimation network is applied to estimate the parameters of the exposure correction model composed of gamma transformation and logistic function from the image, and the model is applied to adjust the response value of the image to realize the correction of the image exposure.

[0005] The technical scheme of the present application is an image exposure correction method based on deep curve estimation, which specifically includes the following steps:

[0006] Step 1: Create a dataset with images of different exposure levels using a camera and image processing software, which specifically includes the following steps:

[0007] Step 1.1, manually collect the optimal exposure image using a camera;

[0008] Step 1.2, render the optimal exposure image and the improperly exposed image using image post-processing software;

[0009] Step 2, build and train a deep curve estimation model, specifically including the following steps:

[0010] Step 2.1, image preprocessing, scale the input image to a size of 256 on the short side, and crop the image to a size of 256x256 from the center;

[0011] Step 2.2, based on a convolutional neural network, build a feature extraction module for extracting image features;

[0012] Step 2.3, combine the feature extraction module with a neural network to form a gamma estimation module for estimating gamma correction parameters;

[0013] Step 2.4, combine the feature extraction module with two neural networks to build a logistic estimation module for estimating logistic function parameters;

[0014] Step 2.5, merge the gamma estimation module and the logistic estimation module to build a deep curve estimation network model;

[0015] Step 2.6, establish a composite loss function and optimize the deep curve estimation model;

[0016] Step 3, estimate the parameters of the image exposure correction curve, correct the exposure of the image using the image exposure correction model composed of gamma transformation and logistic function, specifically including the following steps:

[0017] Step 3.1, estimate the parameters of the image exposure correction curve using the deep curve estimation model;

[0018] Step 3.2, establish an exposure correction model composed of gamma transformation and logistic function to correct the exposure of the image.

[0019] Further, the method for manually collecting the optimal exposure image using a camera in step 1.1 is to select different scenes, manually optimize the exposure settings of the camera and take photos to obtain RAW format images with optimal exposure.

[0020] Further, the different exposure image rendering method in step 1.2 is to apply image post-processing software to render RAW format images into ideal exposure images, and on this basis, adjust the exposure settings in the software to increase or decrease the image exposure, respectively render underexposed and overexposed images, and complete the creation of the data set.

[0021] Further, the specific implementation of the feature extraction module in step 2.2 is as follows:

[0022] The input of the feature extraction module is an image with a size of 256x256x3. The module first extracts a feature map with 16 channels and a size of 256x256 from the input image using a convolution layer with a kernel size of 3x3. Then, the feature extraction module uses three BRC modules to further map the feature map. In each BRC module, there are three basic components (referred to as BRC blocks) composed of batch normalization (BN), ReLU non-linear function and 3x3 convolution. The BRC block does not modify the spatial size of the input feature map, but expands the feature representation of the image by increasing the number of channels of the feature map. The number of channels of the output feature map of each BRC block is 16 more than that of the input feature map.

[0023] To avoid the feature dimension extracted by the feature extraction module at the end being too high, leading to feature dimension explosion, a transition module is used in the first two feature extraction modules to compress the dimension of the intermediate feature. The transition module uses a 1x1 convolution to reduce the channel dimension of the BRC module output feature map to 16, and then uses a mean pooling layer to reduce the spatial dimension of the feature map to one quarter of the original, obtaining a more compact feature representation of the image.

[0024] Further, in step 2.3, the gamma estimation module is composed of the aforementioned feature extraction module and a neural network. The neural network in the gamma estimation module contains a hidden layer with 1000 neurons, followed by a ReLU non-linear layer for non-linear mapping. In training, the response of the neurons in the hidden layer of the neural network is discarded with a probability of 50%. Since the output value of the gamma estimation module is a real number greater than 0, the output layer of the network is only set with one neuron, and a ReLU function is set after it to map the output of the neuron to the non-negative axis.

[0025] Further, in step 2.4, the components of the logistic estimation module include the aforementioned feature extraction module, a neural network for estimating the slope, and a neural network for estimating the midpoint. The neural network for estimating the slope is composed of a hidden layer with 1000 nodes and an output layer with only one node. The hidden layer is followed by a hyperbolic tangent function for non-linear mapping, and the response of the nodes is discarded with a probability of 50% in training. Since there is no constraint on the value of the slope of the logistic function, the output of the neural network is the result of the original linear mapping.

[0026] The neural network for estimating the midpoint is also composed of a hidden layer with 1000 nodes and an output layer with only one node, and a hyperbolic tangent function is used for non-linear mapping after the hidden layer, and the response of the node is discarded with a probability of 50% in training. The output layer is subjected to a Sigmoid non-linear mapping to map the result to the open interval (0, 1).

[0027] Further, in step 2.5, the specific combination of the gamma estimation module and the logistic estimation module is to merge the gamma estimation module and the logistic estimation module in series, if the two modules are represented as E gamma (.) and E logistic (.) respectively, and the input image of the model is represented as I, then the overall operation process of the merged model is:

[0028]

[0029] Where γ is the output of the gamma estimation module, I γ is the luminance compensation result output in the middle of the model, a and k correspond to the midpoint and slope estimated by the logistic estimation module respectively, I out is the output of the model, that is, the final contrast enhancement result completed by the model.

[0030] Further, the composite loss function L all (.) in step 2.6 is composed of a luminance loss term and a contrast loss term, which can be represented as:

[0031]

[0032] Where H and W represent the height and width of the image respectively, i represents the pixel index, and I1 and I2 represent two images; the luminance loss is the error between the luminance compensation result I brightness output in the middle of the model and the real image I gt , and the contrast loss is the error between the image I contrast after contrast enhancement and the real image, and the function for calculating the error of each loss term is a smooth L1 loss function f(x, y), x and y represent two inputs of the function, and the random gradient descent method is used to optimize the deep model.

[0033] Further, in step 3.1, the estimation method of the image exposure correction curve is to scale the image to a size with a short side of 256 pixels, and cut a 256x256 image from the middle of the image as the input of the curve estimation model, and the output of the gamma estimation module in the curve estimation model is used as the gamma parameter in the exposure correction curve, and the output of the logistic estimation module is used as the midpoint and slope of the exposure correction curve.

[0034] Further, the correction model composed of the gamma transformation and the logistic function in step 3.2 is as follows:

[0035]

[0036] Wherein, I represents the original size image, I out represents the correction result, gamma represents the gamma parameter of the curve, a represents the midpoint of the curve, k represents the slope of the curve, and the parameters gamma, a and k are determined through step 3.1.

[0037] The image exposure correction method provided by the application realizes general and accurate image exposure correction, and the subjective and objective comparisons with other methods on the international general data set show that the method can obtain more superior exposure correction effect. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The technical scheme of the application is implemented by the flowchart.

[0039] Figure 2 The deep curve estimation model is constructed according to the embodiment of the application.

[0040] Figure 3 The comparison results of the application and other methods on the overexposed image are shown.

[0041] Figure 4 The comparison results of the application and other methods on the underexposed image are shown. DETAILED DESCRIPTION

[0042] In order to more specifically and clearly show the technical scheme of the application, the implementation of the application will be further described in detail below with reference to the drawings. Although the following embodiments involve many technical details related to the application, these details are only used to illustrate the principles of the application, and those skilled in the art can make appropriate adjustments and changes according to needs without departing from the spirit of the application. In addition, in order to avoid excessive complexity of the content of the embodiments, the prior art used in the embodiments is not described in detail in the subsequent content, but this does not hinder those skilled in the art from understanding the connotation of these technologies, so it will not affect the sufficiency of the disclosure of this specification.

[0043] The application will be further described below with reference to the drawings.

[0044] Figure 1 The flowchart of the embodiment is shown, and according to the flowchart, the specific steps in the embodiment include:

[0045] 1. Exposure correction data set creation: using a camera and image processing software, a data set of images with different exposure is created, which includes the following implementation steps.

[0046] (1.1) Optimal exposure image collection

[0047] A professional photographer uses a digital camera to take pictures of scenes such as cityscapes, landscapes, and people. The camera exposure parameters are manually adjusted during shooting to obtain an ideal exposure image, and the RAW format file of the image is saved.

[0048] (1.2) Rendering images of different exposures

[0049] The image is rendered using the Photo Shop Camera Raw image post-processing software. First, the original ideal exposure image is directly rendered into a dataset real image. Then, using the exposure adjustment tool in the post-processing software, an overexposed image is rendered by increasing the exposure by 1.5 on the basis of the ideal exposure, and an underexposed image is rendered by reducing the exposure by 1.5 on the basis of the ideal exposure, completing the creation of the dataset. Subsequently, the dataset is randomly divided without overlap, with 3535 pairs of images as the training set, 150 pairs of images as the validation set, and 1181 pairs of images as the test set, for subsequent model training and testing.

[0050] 2. Construction and training of depth curve estimation model: based on the convolutional neural network, a depth modeling library Pytorch is applied to construct and train a depth model that can estimate the exposure correction curve. Further, combined with Figure 2 This step includes the following specific implementation steps.

[0051] (2.1) Image preprocessing

[0052] The input image is scaled to a size of 256 on the short side, and the image is cropped to a size of 256x256 from the center.

[0053] (2.2) Construction of feature extraction module

[0054] The feature extraction module takes a 256x256 image as input. First, a 256x256x16 feature map is extracted from the input image using a 3x3 convolutional layer. Then, the feature extraction module uses three BRC modules to further map the feature map. Each BRC module contains three basic components (referred to as BRC blocks) consisting of batch normalization (BN), ReLU non-linear function, and 3x3 convolution. The number of channels of the feature map output by each BRC block is 16 more than the number of channels of the input feature map, and the spatial dimension of the feature is maintained.

[0055] To avoid the feature dimension extracted by the last feature extraction module being too high, leading to feature dimension explosion, a transition module is used in the first two feature extraction modules to compress the dimension of the intermediate feature. In the transition module, a 1x1 convolution is used to reduce the channel dimension of the BRC module output feature map to 16, and then a mean pooling layer is used to reduce the spatial dimension of the feature map to one fourth of the original. After feature flattening, the image is represented as a 16384-dimensional vector.

[0056] (2.3) Construction of Gamma Estimation Module

[0057] The Gamma estimation module is composed of a feature extraction module and a neural network. The hidden layer of the neural network in the Gamma estimation module is provided with 1000 nodes, and is followed by a ReLU nonlinear layer. In training, the response of the neurons in the hidden layer of the neural network is discarded at a probability of 50%. Since the Gamma estimation module needs to output a real number greater than 0, the output layer of the network is only provided with one neuron, and a ReLU function is provided after it to map the output of the neuron to the non-negative axis.

[0058] (2.4) Construction of Logistic Estimation Module

[0059] The Logistic estimation module is composed of a feature extraction module, a neural network for estimating the slope, and a neural network for estimating the midpoint. In the neural network for estimating the slope, a hidden layer with 1000 nodes is provided, and the output layer is only provided with one node. The hidden layer is followed by a hyperbolic tangent function for nonlinear mapping, and the response of the nodes is discarded at a probability of 50% in training. Since the value of the slope of the Logistic function has no constraints, the output of the output layer of the network is the result of the original linear mapping.

[0060] The neural network for estimating the midpoint is also composed of a hidden layer with 1000 nodes and an output layer with only one node, and the hidden layer is followed by a hyperbolic tangent function for nonlinear mapping, and the response of the nodes is discarded at a probability of 50% in training. The output layer will be subjected to a Sigmoid nonlinear mapping to map the output to the open interval (0, 1).

[0061] (2.5) Integration of Modules to Establish Deep Curve Estimation Network Model

[0062] The Gamma estimation module and the Logistic estimation module are combined in series. If the two modules are represented as E gamma (.) and E logistic (.) respectively, and the input image of the model is represented as I, then the overall operation process of the combined model is:

[0063]

[0064] where γ is the output of the gamma estimation module, I γ is the output of the model, a and k are the midpoint and slope estimated by the logistic estimation module, respectively, and I out is the output of the model, i.e., the result of the contrast enhancement completed by the model.

[0065] (2.6) Constructing the composite loss function and training the model

[0066] The composite loss function L all used to train the model is composed of a luminance loss term and a contrast loss term, which can be expressed as:

[0067]

[0068] where H and W represent the height and width of the image, respectively, i represents the pixel index, and I1 and I2 represent two images; the luminance loss is the error between the luminance compensation result I brightness output by the model and the real image I gt , while the contrast loss is the error between the contrast-enhanced image I contrast and the real image, and the function L(.) used to calculate the error of each loss term is the smooth L1 loss function f(x, y), where x and y represent the two inputs of the function.

[0069] The composite loss function is used to calculate the error of the depth curve estimation model on the training set, and the Adam optimizer is used to train the model for two hundred rounds. During training, the initial learning rate is set to 0.0001, and after every 20 rounds of training, the learning rate is halved. After each round of training, the effect of the model is detected on the validation set, and the model with the best effect on the validation set in the two hundred rounds of training is selected as the optimal model.

[0070] 3. Estimate the parameters of the image exposure correction curve using the depth curve estimation model, and correct the exposure of the image using the image exposure correction model composed of the gamma transformation and the logistic function. The specific implementation is as follows.

[0071] (3.1) Estimate the parameters of the image-specific exposure correction curve

[0072] The images in the test set are scaled proportionally to a size of 256 pixels on the short side, and a 256x256 image is cut from the middle of the image as the input of the depth curve estimation model. The output of the gamma estimation module in the curve estimation model is used as the gamma parameter in the exposure correction curve, and the output of the logistic estimation module is used as the midpoint and slope of the exposure correction curve.

[0073] (3.2) Correcting image exposure using exposure correction model of composite gamma transformation and logistic function

[0074] The correction model formed by the composite of gamma transformation and logistic function is as follows:

[0075]

[0076] wherein I represents the original size image, I out represents the correction result, γ represents the gamma parameter of the curve, a represents the midpoint of the curve, and k represents the slope of the curve. The parameters γ, a and k are obtained through step (3.1) and are brought into the model. Finally, the original image is response mapped using the model to correct the image exposure.

[0077] In this embodiment, the mean square error (MSE), the mean angular error (MAE, unit: degree), the color difference (ΔE 00 ), the peak signal noise ratio (PSNR) and the structural similarity (SSIM) are used as error measurement indexes respectively to evaluate the average correction accuracy of the present application on overexposed images and underexposed images on the test set and compare with the prior art. The results of quantitative comparison are shown in Table 1, and the results of subjective comparison are shown in Figure 3 、 Figure 4

[0078] Table 1 Quantitative comparison of the present application and 12 prior art methods

[0079]

[0080]

[0081] The “Auto Levels”, “Auto Contrast”, “Auto Level” and “Auto Level+Contrast” in Table 1 are image exposure adjustment methods provided in Photoshop software; “He”, “CLAHE”, “LIME”, “DeepUPE”, “DUAL”, “RetinexNet”, “DCE” and “Multi-Scale” are the methods proposed in documents [1] to [8] in order.

[0082] ​[1] Zhang Z, Xu C, Ren S, et al. Digital Image Processing and Machine Vision - Visual C++ and MATLAB Implementation (2nd Edition) [M]. People's Posts and Telecommunications Press, 2014: 82-86.

[0083] [2] Zheng L, Yu W, Dong Y. Infrared Image Enhancement Algorithm Based on Dark Channel Prior and CLAHE [J]. Computer Engineering and Design, 2015, 36(12): 3297-3301.

[0084] [3] Guo X, Li Y, Ling H. LIME: Low-Light Image Enhancement via Illumination Map Estimation [J]. IEEE Transactions on Image Processing, 2017, 26(2): 982-993.

[0085] [4] R. W, Q. Z, C. W F, et al. Underexposed Photo Enhancement Using Deep Illumination Estimation [A]. In: 2019 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019 [C]. 15-20 June 2019: 6842-6850. [5] S. M, P. M, S. M, et al. DeepLPF: Deep Local Parametric Filters for Image Enhancement [A]. In: 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020 [C]. 13-19 June 2020: 12823-12832.

[0086] [5] Zhang Q, Nie Y, Zheng W. Dual Illumination Estimation for Robust Exposure Correction [J]. Computer Graphics Forum, 2019, 38(7): 243-252.

[0087] [6] Wei C, Wang W, Yang W, et al. Deep retinex decomposition for low-light enhancement [A]. British Machine Vision Conference, 2018(61772043): 1-12.

[0088] [7] C. G, C. L, J. G, et al. Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement [A]. In: 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020 [C]. 13-19 June 2020: 1777-1786.

[0089] [8] M. A, K. G D, B. O, et al. Learning Multi-Scale Photo Exposure Correction [A]. In: 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021 [C]. 20-25 June 2021: 9153-9163.

[0090] Table 1 and Figure 3 , Figure 4 The results of Table 1 and

[0091] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

Claims

1. An image exposure correction method based on depth curve estimation, characterized in that, Includes the following steps: Step 1 involves acquiring images using a camera and rendering them at different exposure levels using image post-processing software to obtain a dataset of images with varying exposures. This step includes the following steps: Step 1.1: Manually use a camera to capture the optimally exposed image; Step 1.2: Use image post-processing software to render images with different exposure levels; Step 2, construct and train a depth curve estimation model, which includes the following steps: Step 2.1, Image preprocessing; Step 2.2: Based on the convolutional neural network, construct a feature extraction module for extracting image features; Step 2.3: Combine the feature extraction module with a single neural network to construct a gamma estimation module for estimating gamma correction parameters; Step 2.4: Combine the feature extraction module with the two neural networks to construct a logistic estimation module for estimating the parameters of the logistic function; In step 2.4, the logistic function estimation module is constructed as follows; The logistic estimation module is used to estimate the midpoint and slope of the logistic function. Its structure includes: a feature extraction module, a neural network for estimating the slope, and a neural network for estimating the midpoint. The neural network for estimating the slope consists of a hidden layer with 1000 nodes and an output layer with only one node. The hidden layer is followed by a hyperbolic tangent function for nonlinear mapping, and the responses of its nodes are discarded during training with a probability of P%. Since there are no constraints on the value of the logistic function slope, the output of the network output layer is the result of the original linear mapping, where P is a set constant. The neural network for estimating the midpoint also consists of a hidden layer with several nodes and an output layer with only one node. The hidden layer is followed by a hyperbolic tangent function for nonlinear mapping. The node's response is discarded during training with a probability of P%. The output layer is subjected to a sigmoid nonlinear mapping to map the result to the open interval (0, 1). Step 2.5: Merge the gamma estimation module and the logistic estimation module to establish a depth curve estimation model, as follows; The gamma estimation module is connected in series. (.) and logistic estimation module If we combine the input image of the depth curve estimation model and represent it as I, then the overall computational flow of the combined depth curve estimation model is as follows: （1）； in For the output of the gamma estimation module, and These correspond to the midpoint and slope estimated by the logistic estimation module, respectively. This represents the brightness compensation result output from the intermediate part of the depth curve estimation model. This is the output of the depth curve estimation model, which is the final result of the depth curve estimation model to enhance contrast. Step 2.6: Establish the composite loss function and optimize the depth curve estimation model; Step 3: Estimate the parameters of the image exposure correction curve, and correct the image exposure using an image exposure correction model composed of gamma transform and logistic function. This specifically includes the following steps: Step 3.1: Apply the depth curve estimation model to estimate the parameters of the image exposure correction curve; Step 3.2: Establish an image exposure correction model composed of gamma transform and logistic function to correct image exposure.

2. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: The optimal exposure image acquisition method described in step 1.1 is to select different scenes, manually optimize the camera's exposure settings, and then use the camera to take pictures to obtain the optimally exposed RAW format image.

3. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: The rendering method for images with different exposure levels described in step 1.2 is as follows: using image post-processing software to render the optimal exposure image in RAW format into an ideal exposure image, and then adjusting the exposure settings within the software to increase or decrease the image exposure to render underexposed and overexposed images, thus completing the creation of the dataset.

4. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: The specific implementation method of the feature extraction module in step 2.2 is as follows; The input to the feature extraction module is an image of size 256×256×3. The module first extracts a 256×256×16 feature map from the input image using a 3×3 convolutional layer. Subsequently, the feature extraction module uses three BRC modules to further map the feature map. Each BRC module contains three basic components consisting of batch normalization (BN), a ReLU nonlinear function, and a 3×3 convolution, referred to simply as a BRC block. The BRC block does not modify the spatial size of the input feature map; instead, it expands the image's feature representation by increasing the number of channels in the feature map. Each BRC block outputs a feature map with 16 more channels than the input feature map. To avoid excessively high feature dimensions in the final feature extraction module, leading to feature dimension explosion, a transition module is used in the first two feature extraction modules to compress the dimensions of intermediate features. In the transition module, a 1×1 convolution is used to reduce the channel dimension of the feature map output by the BRC module to 16, and then an average pooling layer is used to reduce the spatial dimension of the feature map to one-quarter of its original size, resulting in a more concise feature representation of the image.

5. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: In step 2.3, the gamma estimation module is constructed as follows; The gamma estimation module consists of a feature extraction module and a neural network. The neural network in the gamma estimation module contains a hidden layer with several neurons, followed by a ReLU nonlinear layer for nonlinear mapping. To avoid overfitting, during training, the responses of neurons in the hidden layer are discarded with a probability of P%. Since the output value of the gamma estimation module is a real number greater than 0, the output layer of this network has only one neuron, and a ReLU function is set after it to map the neuron's output to a non-negative axis, where P is a constant.

6. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: The specific construction method of the composite loss function in step 2.6 is as follows; Composite loss function Composed of a luminance loss term and a contrast loss term, it can be expressed as: （2）； Where H and W represent the height and width of the image, respectively, and i represents the pixel index. and This represents two images; the brightness loss is the brightness compensation result output by the depth curve estimation model. With real images The error between the two sides, while the loss of contrast is a result of contrast enhancement. The error between the image and the real image, and the function for calculating the loss for each loss term. For smoothing L1 loss function, i.e. , and These represent the two inputs to the function; the composite loss function is applied to calculate the error of the depth curve estimation model on the training set, and the stochastic gradient descent method is used to optimize the depth curve estimation model.

7. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: Step 3.1 The method for estimating the image exposure correction curve is as follows: The image is scaled proportionally to a short side of 256 pixels, and a 256×256 image is cropped from the middle of the image as the input to the depth curve estimation model. The output of the gamma estimation module in the depth curve estimation model is used as the gamma parameter in the exposure correction curve, and the output of the logistic estimation module is used as the midpoint and slope of the exposure correction curve.

8. The image exposure correction method based on depth curve estimation as described in claim 1, characterized in that: The image exposure correction model composed of the gamma transform and logistic function in step 3.2 is as follows: （3）； in, Represents the original size of the image. Indicates the correction result. The gamma parameter represents the curve. This represents the parameter of the midpoint of the curve. Indicates the slope of the curve, parameter , , All were obtained through step 3.1.

Images