Unsupervised deep learning sand and dust image color correction method based on color cast physical model

Abstract

The invention discloses an unsupervised deep learning sand and dust image color correction method based on a color cast physical model, and the method comprises the steps: obtaining a first image with sand and dust and a second image without sand and dust in a same scene, and reconstructing the first image and the second image through a first generator and a second generator, the method comprises the following steps: reconstructing a predicted image with sand and dust to obtain a reconstructed predicted image without sand and dust and a predicted image with sand and dust, respectively calculating corresponding losses according to the images with sand and dust and the images without sand and dust, fusing the losses to obtain a target loss, obtaining a first generator after fine tuning according to the target loss, and obtaining a predicted image with sand and dust according to the first generator after fine tuning. And performing dust removal reconstruction on the first image to obtain a reconstructed dust removal image. According to the invention, the prediction image generated by the generator and the real image are alternately discriminated and calculated, reliable support is provided for fine adjustment of the generator, the generated prediction image is closer to the real image, and thus the dust removal reconstruction effect of the dust image is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing and deep learning technology, and in particular to an unsupervised deep learning dust image color correction method based on a color cast physical model. Background Art

[0002] Dust storm is a common bad weather. When shooting in this environment, the light beam is affected by sand and dust particles, which causes attenuation and scattering, and eventually leads to reduced image quality, which is called dust-degraded image. Most shooting systems, such as automatic navigation and intelligent monitoring, require the input image to be clear and objective. Therefore, the enhancement technology of dust-degraded images is an important research topic in the field of computer vision, and its application prospects are very broad.

[0003] Generally, the two most obvious disadvantages of dust-degraded images are color distortion and reduced contrast. If the color distortion problem can be completely solved, the dust-degraded image will be similar to a foggy image, and the foggy image enhancement method can be used to restore the contrast. Therefore, color correction becomes the key point of dust-degraded image enhancement.

[0004] In most methods of enhancing images degraded by dust, the grayscale world method is usually used to correct the color of the image. For example, the application number is: CN202111180437, and the invention name is: A method for enhancing dust images, which borrows the idea of ​​the grayscale world method, calculates the gain according to the mean of the three channels, and stretches the mean of the three channels to the same value. However, the correction model of this method is too simple and has poor generalization ability; in CN118447128A, and the invention name is: A method, device, equipment and medium for dust removal and reconstruction of dust images, color correction is completed by training a deep learning network, but in this technology, it is necessary to obtain a first image with dust and a second image without dust in the same scene, and this step is very difficult.

[0005] In recent years, the powerful fitting ability of deep learning models has been fruitful in various fields of image processing. However, deep learning training usually requires a large amount of paired data. It is very difficult to take images of the same scene with and without sandstorms, which makes it difficult to train relevant deep learning models. Summary of the invention

[0006] The purpose of the present invention is to solve the defects in the above-mentioned prior art and the technical problems are embodied in the following points:

[0007] 1) Combined with the color cast physical model, an accurate and generalized deep learning model is provided for correcting the color of dust images; 2) The training of the deep learning model is completed without the need for paired data sets, solving the problem of difficulty in collecting data sets. The present invention overcomes the problem that supervised learning networks require a large amount of manual annotation while also overcoming the problem that traditional algorithms are not very applicable. The specific technical solution to achieve the purpose of the present invention is:

[0008] The technical solution adopted by the present invention is based on the GAN model and combines the physical model of color cast of dust images to propose an unsupervised deep learning dust image color correction method based on the color cast physical model, which includes the following steps:

[0009] Step S1: Obtain an atmospheric image dataset and a dust image dataset. For atmospheric images, collect them from annotated and public datasets on the Internet; for dust images, collect them from the SIE dataset and papers that use traditional algorithms to enhance dust images.

[0010] Step S2: construct an adversarial generative network model based on the GAN model, which includes a generator G and a discriminator D. The generator G is used to generate forged dust image data, and the discriminator D is used to detect whether the image generated by the generator G is close to the real dust image data; input the atmospheric image dataset and the dust image dataset into the constructed network model, calculate the loss through the loss function during the training process, continuously iterate and update the model parameters, and finally obtain the trained generator and discriminator; then input the atmospheric image dataset into the generator to obtain the forged dust image data in the corresponding scene; so far, a paired dataset of dust color cast images and atmospheric non-color cast images has been generated;

[0011] Step S3: Construct a deep learning network model for color correction. Through the training process, the model can accurately correct the color-cast images affected by dust. The pre-prepared paired data set is input into the network. Through this process, the network will learn how to identify and correct the color deviation from the color-cast images, and finally achieve effective color correction of the images affected by dust. Further, the step S2 specifically includes:

[0012] S2.1: Build the generator

[0013] S2.1.1: Input the atmospheric image I1 and its corresponding depth map R1, where I1 contains three color channels R, G, B, and R1 has only one channel; superimpose I1 and R1 to obtain a 4-channel feature map G0;

[0014] S2.1.2: G0 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G1.

[0015] S2.1.3: G1 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G2.

[0016] S2.1.4: Perform matrix dot multiplication on G2 and I1 to obtain a new 3-channel feature map G3;

[0017] Steps S2.1.2 to S2.1.4 are used to learn the color attenuation process G3 = I1·e in a dusty environment -R1·N , where R1 is the depth map and N represents the attenuation coefficient;

[0018] S2.1.5: Input the randomly generated signal Z, which passes through the fully connected layer, a convolution layer with a kernel size of 3, a stride of 1 and a channel number of 1, and then passes through the BN layer and the activation function ReLU to obtain the feature map Z1 with a channel number of 1;

[0019] S2.1.6: Superimpose Z1 and R1 to obtain a feature map G4 with 2 channels. Pass G4 through a convolution layer with a kernel size of 3, a stride of 1, and 8 channels. The resulting feature map is called G5.

[0020] S2.1.7: Pass G5 through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3 to obtain a feature map G6 of the same size as G1;

[0021] S2.1.8: Add the obtained feature maps, G7 = G1 + G6; It is generally believed that when light propagates in a dusty environment, it will experience a scattering effect, and these scattered lights will return to the image sensor, thus forming a hazy effect in the dusty image. Therefore, steps S2.1.5 to S2.1.8 learn the scattering process of colors in a dusty environment by superimposing a mask on the original image;

[0022] S2.1.9: G7 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G8.

[0023] S2.1.10: G8 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G9.

[0024] S2.1.11: G9 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G_fake; G_fake is the fake dust image generated by the generator; S2.2: Construct a discriminator

[0025] S2.2.1: Input a three-channel image, called I2; scale the size of I2 to 256*256;

[0026] S2.2.2: I2 passes through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 16, and then passes through a BN layer and an activation function ReLU. The resulting feature map is called D1, and the size of D1 is 128*128;

[0027] S2.2.3: Pass D1 through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 64, and then through a BN layer. The resulting feature map is called D2, and the size of D2 is 64*64;

[0028] S2.2.4: D2 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 64, and then through a BN layer and an activation function ReLU. The resulting feature map is called D3, and the size of D3 is 64*64.

[0029] S2.2.5: D3 is passed through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 128, and then through a BN layer. The resulting feature map is called D4, and the size of D4 is 32*32.

[0030] S2.2.6: D4 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 128, and then through a BN layer and an activation function ReLU. The resulting feature map is called D5, and the size of D5 is 32*32.

[0031] S2.2.7: D5 is passed through a fully connected layer of size 32*32*128 and activated by the Softmax function. The resulting value is called D_out. D_out is the credibility score of the discriminator for whether the image is a dust image.

[0032] S2.3: Training Model

[0033] The loss function is used to optimize the generator and discriminator models; the loss function is adversarial loss 3. The unsupervised deep learning dust image color correction method according to claim 1, characterized in that step S3 specifically comprises:

[0034] S3.1: Input a fake dust image, called I3; I3 contains three channels: R, G, and B;

[0035] S3.2: The three channels are passed through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature maps are called R1, Gr1, and B1 respectively;

[0036] S3.3: The obtained R1, Gr1 and B1 feature maps are respectively passed through a convolution layer with a convolution kernel size of 3, a stride of 1 and a channel number of 1, and then through a BN layer and an activation function ReLU. The obtained feature maps are called R2, Gr2, and B2 respectively;

[0037] S3.4: Superimpose R2, Gr2, and B2 to obtain a feature map called F1; pass F1 through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU to obtain a feature map called F2;

[0038] S3.5: F2 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called F3.

[0039] S3.6: Pass F3 through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 1, and then through a BN layer and an activation function Softmax. The resulting feature map is called F_out, which is the model after color correction.

[0040] S3.7: Calculate the L2 loss function of F_out and the paired real atmospheric image y: Loss = (F_out-y) 2 , where the smaller the Loss, the better the fitting effect; and the network parameters are trained based on this to obtain a network model with color correction function.

[0041] Due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0042] 1) A dust image color correction model and dataset generation algorithm based on deep learning and without the need for paired data.

[0043] 2) A generator structure designed for the physical model of dust image color cast is used to forge dust images paired with atmospheric images, making the model training process more stable and capable of generating a dataset that is indistinguishable from real dust images.

[0044] 3) The optimized color correction model structure first extracts features from each channel separately so that the model can better calculate the color distribution within a single channel, and then perform color correction on this basis, which can better and stably correct the color cast problem of dust images.

[0045] 4) The overall structure of the model is lightweight and easy to reproduce and deploy in a real-time shooting system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of the present invention;

[0047] Figure 2 The present invention constructs a flow chart of a generator model;

[0048] Figure 3 The present invention constructs a discriminator model flow chart;

[0049] Figure 4 The present invention constructs a color correction model flow chart;

[0050] Figure 5 A comparison chart of the color correction results of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] See also Figure 1 The present invention provides an unsupervised deep learning dust image color correction method based on a color cast physical model, comprising the following steps:

[0053] Step S1: Obtain atmospheric image datasets and dust image datasets. For atmospheric images, there are currently a large number of annotated and public datasets on the Internet, such as "Microsoft 7-Scenes", "NYU Depth", "UW RGB-DObject", "B3DO", etc. These datasets contain images in air environments and corresponding depth maps, which can be used for subsequent training. For dust image datasets, they are collected from the "SIE" dataset and some papers that use traditional algorithms for dust image enhancement.

[0054] Step S2: Construct an adversarial generative network model based on the GAN model, which includes a generator G and a discriminator D. The generator G is used to generate forged dust image data, and the discriminator D is used to detect whether the image generated by the generator G is close to the real dust image data. Input the atmospheric image dataset and the dust image dataset into the constructed network model, calculate the loss through the loss function during the training process, continuously iterate and update the model parameters, and finally obtain the trained generator and discriminator. Then input the atmospheric image dataset into the generator to obtain the forged dust image data in the corresponding scene. So far, a paired dataset of dust color cast images and atmospheric non-color cast images has been generated.

[0055] Step S3: construct a color correction network model, input the generated paired data set into the network model, and obtain a deep learning network that can correct the color cast of dust images through training.

[0056] Furthermore, step S2 includes three steps: constructing a generator S2-1, constructing a discriminator S2-2, and training a model S2-3.

[0057] The construction of S2-1 specifically includes the following:

[0058] S2-1-1: Input the atmospheric image I1 and its corresponding depth map R1, where I1 contains three color channels: R, G, and B, and R1 has only one channel. I1 and R1 are superimposed to obtain a 4-channel feature map G0.

[0059] S2-1-2: G0 passes through a convolution layer with a kernel size of 3, a step size of 1, and 8 channels, and then passes through a BN layer and an activation function ReLU. The resulting feature map is called G1.

[0060] S2-1-3: G1 is passed through a convolution layer with a kernel size of 3, a step size of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G2.

[0061] S2-1-4: Perform matrix dot multiplication on G2 and I1 to obtain a new 3-channel feature G3. Steps S2-1-2 to S2-1-4 are used to learn the color attenuation process in a dusty environment G3 = I1·e -R1·N .

[0062] S2-1-5: Input the randomly generated signal Z, which passes through the fully connected layer, a convolution layer with a convolution kernel size of 3, a step size of 1, and a channel number of 1, and passes through the BN layer and the activation function ReLU to obtain the feature map Z1 with a channel number of 1.

[0063] S2-1-6: Superimpose Z1 and R1 to obtain a feature map G4 with a channel number of 2. Pass G4 through a convolution layer with a convolution kernel size of 3, a step size of 1, and a channel number of 8. The resulting feature map is called G5.

[0064] S2-1-7: Pass G5 through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3 to obtain a feature map G6 of the same size as G1.

[0065] S2-1-8: Add the feature maps obtained above, Steps S2-1-5 to S2-1-8 are used to learn the color scattering process in a dusty environment.

[0066] S2-1-9: G7 is passed through a convolution layer with a kernel size of 3, a stride of 1, and 8 channels, and then through a BN layer and an activation function ReLU. The resulting feature map is called G8.

[0067] S2-1-10: G8 is passed through a convolution layer with a kernel size of 3, a stride of 1, and 8 channels, and then through a BN layer and an activation function ReLU. The resulting feature map is called G9.

[0068] S2-1-11: G9 is passed through a convolution layer with a kernel size of 3, a step size of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G_fake. G_fake is the fake dust image generated by the generator. The construction of S2-2 specifically includes the following:

[0069] S2-2-1: Input a three-channel image, called I2. Scale the size of I2 to 256*256. Since the discriminator is only used to judge the color style of the image, it pays less attention to the details and texture of the image, so the input image can be scaled. S2-2-2: I2 passes through a convolution layer with a convolution kernel size of 3, a stride of 2, and 16 channels, and passes through a BN layer and an activation function ReLU. The resulting feature map is called D1, and the size of D1 is 128*128.

[0070] S2-2-3: D1 passes through a convolution layer with a kernel size of 3, a step size of 2, and 64 channels, and then passes through a BN layer. The resulting feature map is called D2, and the size of D2 is 64*64.

[0071] S2-2-4: D2 passes through a convolution layer with a kernel size of 3, a step size of 1, and 64 channels, and then passes through a BN layer and an activation function ReLU. The resulting feature map is called D3, and the size of D3 is 64*64.

[0072] S2-2-5: D3 passes through a convolution layer with a kernel size of 3, a step size of 2, and a channel number of 128, and then passes through a BN layer. The resulting feature map is called D4, and the size of D4 is 32*32.

[0073] S2-2-6: D4 is passed through a convolution layer with a kernel size of 3, a step size of 1, and a channel number of 128, and then through a BN layer and an activation function ReLU. The resulting feature map is called D5, and the size of D5 is 32*32.

[0074] S2-2-7: D5 is passed through a fully connected layer of size 32*32*128 and activated by the Softmax function. The resulting value is called D_out. D_out is the credibility score of the discriminator for whether the image is a dust image.

[0075] In the training step S2-3, the loss function is used to optimize the generator and discriminator models. The loss function is adversarial loss This step uses a more classic loss function, so I will not go into details.

[0076] The construction of S3 specifically includes the following:

[0077] S3-1: Input a fake dust image, called I3. I3 contains 3 channels: R, G, B.

[0078] S3-2: The three channels are passed through a convolution layer with a kernel size of 3, a step size of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature maps are called R1, Gr1, and B1 respectively.

[0079] S3-3: The three newly obtained feature maps are respectively passed through a convolution layer with a convolution kernel size of 3, a step size of 1, and a channel number of 1, and then through a BN layer and an activation function ReLU. The obtained feature maps are called R2, Gr2, and B2 respectively.

[0080] S3-4: Superimpose R2, Gr2, and B2 to obtain a new feature map, called F1. Pass F1 through a convolution layer with a kernel size of 3, a step size of 1, and 8 channels, and pass it through a BN layer and an activation function ReLU. The feature maps obtained are called F2. S3-5: Pass F2 through a convolution layer with a kernel size of 3, a step size of 1, and 8 channels, and pass it through a BN layer and an activation function ReLU. The feature maps obtained are called F3.

[0081] S3-6: F3 is passed through a convolution layer with a kernel size of 3, a step size of 1, and a channel number of 1, and then through a BN layer and an activation function Softmax. The obtained feature maps are called F_out. F_out is the model after color correction.

[0082] S3-7: Calculate the L2 loss function between F_out and the paired real atmospheric image, and use it to train the network parameters to obtain a network model with color correction function.

[0083] Example 1

[0084] Step 1: Obtain a dust image dataset and a color-free atmospheric image dataset with depth information.

[0085] Step 2: Build the generator model G and the discriminator model D.

[0086] Step 3: Input the color-free atmospheric image I1 and its corresponding depth map R1 into the generator model, as Figure 2 , output the fake dust image G_fake.

[0087] Step 4: Input the fake dust image G_fake into the discriminator model, such as Figure 3 , output credibility F_out.

[0088] Step 6: Calculate the loss function LOSS1 = log(1-F_out1).

[0089] Step 7: Input the real dust image into the discriminator model and output the credibility F_out2.

[0090] Step 8: Calculate the loss function LOSS2 = log(F_out1).

[0091] Step 9: Calculate the total loss LOSS = LOSS1 + LOSS2, and optimize the network parameters of the discriminator D based on the result.

[0092] Step 10: Input the color-free atmospheric image I1 and its corresponding depth map R1 into the generator model, and output the fake dust image G_fake.

[0093] Step 11: Input the fake dust image G_fake into the discriminator model and output the credibility F_out.

[0094] Step 12: Calculate the loss function LOSS = log(F_out) and optimize the network parameters of the generator G based on the result.

[0095] Step 13: Repeat steps 3 to 12 until the generator G and the discriminator D converge.

[0096] Step 14: Input the color-free atmospheric image dataset and its corresponding depth map into the trained generator model, and output a paired dust image.

[0097] Step 15: Input the dust image output by the generator into the color correction model, such as Figure 4 , and output the corrected image F_out. Step 16, calculate the loss function LOSS = ||F_out-I1|| 2 , and optimize the network parameters of the color correction model according to the results until convergence.

[0098] Step 17: Use the trained color correction model to correct the color cast of the dust image. Figure 5 .

Claims

1. An unsupervised deep learning dust image color correction method based on a color cast physical model, characterized in that: The following steps are involved: Step S1: Acquire an atmospheric image dataset and a dust image dataset. For atmospheric images, collect them from annotated and public datasets on the Internet. For dust images, we collected them from the SIE dataset and papers that used traditional algorithms for dust image enhancement; Step S2: construct an adversarial generative network model based on the GAN model, which includes a generator G and a discriminator D. The generator G is used to generate forged dust image data, and the discriminator D is used to detect whether the image generated by the generator G is close to the real dust image data; input the atmospheric image dataset and the dust image dataset into the constructed network model, calculate the loss through the loss function during the training process, continuously iterate and update the model parameters, and finally obtain the trained generator and discriminator; then input the atmospheric image dataset into the generator to obtain the forged dust image data in the corresponding scene; so far, a paired dataset of dust color cast images and atmospheric non-color cast images has been generated; Step S3: Construct a deep learning network model for color correction. Through the training process, the model can accurately correct the color-cast images affected by sand and dust. Input the pre-prepared paired data set into the network. Through this process, the network will learn how to identify and correct color deviations from color-cast images, and finally achieve effective correction of the color of images affected by sand and dust.

2. The unsupervised deep learning dust image color correction method according to claim 1, characterized in that: The step S2 specifically includes: S2.1: Build the generator S2.1.1: Input the atmospheric image I1 and its corresponding depth map R1, where I1 contains three color channels R, G, B, and R1 has only one channel; superimpose I1 and R1 to obtain a 4-channel feature map G0; S2.1.2: G0 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G1. S2.1.3: G1 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G2. S2.1.4: Perform matrix dot multiplication on G2 and I1 to obtain a new 3-channel feature map G3; Steps S2.1.2 to S2.1.4 are used to learn the color attenuation process G3 = I1·e in a dusty environment -R1·N , where R1 is the depth map and N represents the attenuation coefficient; S2.1.5: Input the randomly generated signal Z, which passes through the fully connected layer, a convolution layer with a kernel size of 3, a stride of 1 and a channel number of 1, and passes through the BN layer and the activation function ReLU to obtain a feature map Z1 with a channel number of 1; S2.1.6: Superimpose Z1 and R1 to obtain a feature map G4 with a channel number of 2, and pass G4 through a convolution layer with a kernel size of 3, a stride of 1 and a channel number of 8. The resulting feature map is called G5; S2.1.7: Pass G5 through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3 to obtain a feature map G6 of the same size as G1; S2.1.8: Add the obtained feature maps, G7 = G1 + G6; Steps S2.1.5 to S2.1.8 learn the color scattering process in the dusty environment by superimposing a mask on the original image; S2.1.9: G7 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G8. S2.1.10: G8 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called G9. S2.1.11: G9 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 3, and then through a BN layer and an activation function Softmax. The resulting feature map is called G_fake; G_fake is the fake dust image generated by the generator; S2.2.1: Input a three-channel image, called I2; scale the size of I2 to 256*256; S2.2.2: I2 passes through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 16, and then passes through a BN layer and an activation function ReLU. The resulting feature map is called D1, and the size of D1 is 128*128; S2.2.3: Pass D1 through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 64, and then through a BN layer. The resulting feature map is called D2, and the size of D2 is 64*64; S2.2.4: D2 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 64, and then through a BN layer and an activation function ReLU. The resulting feature map is called D3, and the size of D3 is 64*64. S2.2.5: D3 is passed through a convolutional layer with a kernel size of 3, a stride of 2, and a channel number of 128, and then through a BN layer. The resulting feature map is called D4, and the size of D4 is 32*32. S2.2.6: D4 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 128, and then through a BN layer and an activation function ReLU. The resulting feature map is called D5, and the size of D5 is 32*32. S2.2.7: D5 is passed through a fully connected layer of size 32*32*128 and activated by the Softmax function. The resulting value is called D_out. D_out is the credibility score of the discriminator for whether the image is a dust image. S2.3: Training Model The loss function is used to optimize the generator and discriminator models; the loss function is adversarial loss 3. The unsupervised deep learning dust image color correction method according to claim 1, characterized in that: The step S3 specifically includes: S3.1: Input a fake dust image, called I3; I3 contains three channels: R, G, and B; S3.2: The three channels are passed through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature maps are called R1, Gr1, and B1 respectively; S3.3: The obtained R1, Gr1 and B1 feature maps are respectively passed through a convolution layer with a convolution kernel size of 3, a stride of 1 and a channel number of 1, and then through a BN layer and an activation function ReLU. The obtained feature maps are called R2, Gr2, and B2 respectively; S3.4: Superimpose R2, Gr2, and B2 to obtain a feature map called F1; pass F1 through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU to obtain a feature map called F2; S3.5: F2 is passed through a convolutional layer with a kernel size of 3, a stride of 1, and a channel number of 8, and then through a BN layer and an activation function ReLU. The resulting feature map is called F3. S3.6: Pass F3 through a convolution layer with a kernel size of 3, a stride of 1, and a channel number of 1, and then through a BN layer and an activation function Softmax. The resulting feature map is called F_out, which is the model after color correction. S3.7: Calculate the L2 loss function of F_out and the paired real atmospheric image y: Loss = (F_out-y) 2 , where the smaller the Loss, the better the fitting effect; and the network parameters are trained based on this to obtain a network model with color correction function.

Images