CN111161159B - Image defogging method and device based on combination of priori knowledge and deep learning - Google Patents

Abstract

The invention relates to an image defogging method and device based on combination of priori knowledge and deep learning. The method comprises the following steps: decomposing the input original foggy image Z by using a weighted guide filter to obtain a detail layer image Z _e And base layer image Z _b (ii) a Using a quadtree search method to search the base layer image Z _b Processing to obtain a global atmosphere light component image A _c Wherein c belongs to { r, g, b }; constructing a deep convolutional neural network on the base layer image Z _b Processing is carried out to obtain a transmissivity image t; utilizing the global atmospheric light component image A based on an atmospheric scattering model _c The transmittance image t and the detail layer image Z _e And restoring the original foggy image Z to obtain a defogged image J.

Description

Image defogging method and device based on the combination of prior knowledge and deep learning

Technical Field

The present invention relates to the field of image processing technology, and in particular to an image defogging method and device based on the combination of prior knowledge and deep learning.

Background Art

In foggy weather, floating particles in the sun, fog or dust cause the image to fade and blur, and the contrast and softness are reduced, which seriously restricts the image quality and limits applications such as video surveillance and analysis, target recognition, urban transportation, aerial photography, and military defense. Therefore, the clarity of foggy images is directly related to people's livelihood and has great practical significance for people's production and life.

At present, dehazing algorithms are mainly divided into three categories: non-model-based dehazing algorithms, model-based dehazing algorithms, and deep learning-based dehazing algorithms. Non-model-based dehazing algorithms mainly enhance the image by directly stretching the contrast of the image. Commonly used methods include: histogram equalization, homomorphic filtering algorithm, algorithm based on Retinex model, and algorithm based on improved Retinex model. These methods perform dehazing based on the principle of optical imaging, making the contrast between image colors more balanced and the image effect softer, but the resulting image cannot be effectively enhanced in contrast. This method weakens the dark or bright areas in the original image and blurs the focus of the image.

Summary of the invention

In view of the deficiencies in the prior art, the present invention provides an image defogging method and device based on a combination of prior knowledge and deep learning.

In a first aspect, the present invention provides an image defogging method based on a priori knowledge combined with deep learning, the method comprising:

Decompose the input original foggy image Z using a weighted guided filter to obtain a detail layer image _Ze and a base layer image _Zb ;

The base layer image Z _b is processed by a quadtree search method to obtain a global atmospheric light component image A _c , where c∈{r,g,b};

Constructing a deep convolutional neural network to process the base layer image Z _b to obtain a transmittance image t;

Based on the atmospheric scattering model, the original foggy image Z is restored using the global atmospheric light component image A _c , the transmittance image t and the detail layer image _Ze to obtain a defogged image J.

Preferably, the input original foggy image Z is decomposed by using a weighted guided filter to obtain a base layer image Z _b , which specifically includes the following steps:

Get the pixel gray value Z(x,y) at any point (x,y) in the original foggy image Z;

The weighted filter coefficients a _p (x, y) and _bp (x, y) at the midpoint (x, y) of the original foggy image Z are obtained using the following formula:

是在原始有雾图像Z中以点(x，y)为中心，ρ为半径的窗体的像素灰度值的方差；ΓY(x,y)是原始有雾图像Z在点(x，y)处的亮度分量，Y＝max{Z_r,Z_g,Z_b}，Z_r,Z_g,Z_b分别是有雾图像Z中点(x，y)处的R,G,B值；λ为大于1的常数；Wherein, μ _Z,ρ (x, y) is the mean value of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z;

is the variance of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z; ΓY(x, y) is the brightness component of the original foggy image Z at point (x, y), Y = max{ _Zr , _Zg , _Zb }, _Zr , _Zg , _Zb are the R, G, B values at point (x, y) in the foggy image Z respectively; λ is a constant greater than 1;

The base layer image Z _b is obtained using the following formula:

Z _b (x, y) = a _p (x, y) Z (x, y) + b _p (x, y);

Wherein, Z _b (x, y) is the grayscale value of the pixel at the point (x, y) in the base layer image Z _b .

Preferably, the method of using a quadtree search method to process the base layer image Z _b to obtain a global atmospheric light component image A _c specifically includes the following steps:

Step A, divide the base layer image Z _b into four rectangular regions Z _bi (i∈{1,2,3,4}), the length and width of Z _bi are 1/2 of the length and width of Z _b respectively;

Step B, defining the score of each rectangular area as the mean value of the pixel grayscale values minus the standard deviation of the pixel grayscale values in the area;

Step C, selecting the rectangular area with the highest score and further dividing it into four rectangular areas;

Step D, repeating steps B and C and iteratively updating the rectangular area with the highest score n times to obtain the final subdivided area Z _b-end ;

The atmospheric light component image A _c is obtained using the following formula:

|A _c (c∈{r,g,b})|=min|((Z _b-end-r (x,y),Z _b-end-g (x,y),Z _b-end-b (x,y))-(255,255,255))|,

Wherein (Z _b-end-r (x, y), Z _b-end-g (x, y), Z _b-end-b (x, y)) is the color vector at the midpoint (x, y) of the final subdivision area Z _b-end , and (255, 255, 255) is a pure white vector.

Preferably, the construction of a deep convolutional neural network to process the base layer image Z _b to obtain the transmittance image t specifically includes the following steps:

The base layer image Z _b is downsampled by using the first convolutional layer to obtain a low-resolution base layer image Z _b ' and then extract low-level features, wherein the formula for downsampling the base layer image Z _b by using the first convolutional layer is:

为所述第一卷积层中第i层卷积层在通道c的索引下，所述低分辨率基础层图像Z_b’的低层特征；x为基础层图像Z_b的横坐标，y为基础层图像Z_b的纵坐标；x'为低分辨率基础层图像Z_b’的横坐标，y'为低分辨率基础层图像Z_b’的纵坐标；

为所述第一卷积层中第i层卷积层在c和c'图层索引通道下的卷积权重数组；

为所述第一卷积层中第i层卷积层在c图层索引通道下的偏差向量；σ(·)＝max(·,0)表示ReLU激活函数，并使用零填充作为所述第一卷积层中所有层卷积层的边界条件；s为所述第一卷积层的卷积核的步长；in,

is the low-level feature of the low-resolution base layer image Z _b ′ under the index of channel c of the i-th convolution layer in the first convolution layer; x is the abscissa of the base layer image Z _b , and y is the ordinate of the base layer image Z _b ; x' is the abscissa of the low-resolution base layer image Z _b ′, and y' is the ordinate of the low-resolution base layer image Z _b ′;

is the convolution weight array of the i-th convolution layer in the first convolution layer under the c and c' layer index channels;

is the bias vector of the i-th convolutional layer in the first convolutional layer under the c-th layer index channel; σ(·)＝max(·,0) represents the ReLU activation function, and zero padding is used as the boundary condition of all convolutional layers in the first convolutional layer; s is the step size of the convolution kernel of the first convolutional layer;

输入层数为n_L＝2的第二卷积层提取局部特征L，其中第二卷积层中卷积核的尺寸为3×3，步长为1；The obtained low-level features

The second convolutional layer with the input layer number n _L = 2 extracts the local feature L, where the size of the convolution kernel in the second convolutional layer is 3×3 and the stride is 1;

输入层数为n_G1＝2、卷积核的尺寸为3×3和步长为2的第三卷积层后，再输入层数为n_G2＝2的全连接层，得到全局特征G；The low-level features obtained by the second convolutional layer

After inputting the third convolutional layer with n _G1 = 2 layers, a convolution kernel size of 3×3 and a stride of 2, the fully connected layer with n _G2 = 2 layers is input to obtain the global feature G;

对应的混合特征图F_L＝σ(L+G)；Add the local feature L and the global feature G and input them into the activation function to get the low-level feature

The corresponding mixed feature map F _L =σ(L+G);

对应的初步大气折射率特征图，然后再对所述初步大气折射率特征图进行上采样，输出得到透射率图像t。Convolution is performed on the mixed feature map F _L =σ(L+G) to obtain the low-level features

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to output a transmittance image t.

Preferably, based on the atmospheric scattering model, the global atmospheric light component image A _c , the transmittance image t and the detail layer image _Ze are used to restore the original foggy image Z, and the formula for obtaining the defogging image J is:

其中η为大于零的常数。Wherein, J(x,y) is the pixel grayscale value at the midpoint (x, y) of the dehazed image J, _Ze (x,y) is the pixel grayscale value at the midpoint (x, y) of the detail layer image _Ze , _Zb (x,y) is the pixel grayscale value at the midpoint (x, y) of the base layer image _Zb , t(x,y) is the transmittance at the midpoint (x, y) of the transmittance image t,

Where η is a constant greater than zero.

对应的初步大气折射率特征图，然后再对所述大气折射率特征图进行上采样，输出得到透射率图像t之后还包括如下步骤：Preferably, the mixed feature map F _L =σ(L+G) is convolved to obtain the low-level feature

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to obtain the transmittance image t, and then the following steps are included:

The loss function is constructed as follows:

L＝ _{Lr + wcLc}

L_c表示为

Among them, _Lr is the reconstruction loss function, _Lc is the color loss function, _wc is the weight assigned to the color loss function _Lc , and _Lr is expressed as

L _c is expressed as

J is the defogging image, Z is the foggy image, c∈(R,G,B) is the channel index, ∠(J(x,y),Z(x,y)) represents the angle between the three-dimensional color vectors of the foggy image and the defogging image at the pixel point (x,y);

The loss function is used to adjust the parameters of the deep convolutional neural network.

In a second aspect, the present invention provides an image defogging device based on a priori knowledge combined with deep learning, the device comprising a memory and a processor;

The memory is used to store computer programs;

The processor is used to implement the image defogging method based on the combination of prior knowledge and deep learning as described above when executing the computer program.

The beneficial effect of the image defogging method and device based on the combination of prior knowledge and deep learning provided by the present invention is that the foggy image is decomposed by a weighted guided filter to obtain a detail layer image and a base layer image, and then the base layer image is processed by a quadtree search method and a deep neural network to obtain a global atmospheric light component image and a transmittance image, and finally the foggy image is restored by using the global atmospheric light component image, the transmittance image and the detail layer image to obtain a defogged image. Since both the global atmospheric light component image and the transmittance image use the base layer image for estimation or calculation, the amplification of image noise is avoided, and the foggy image can be defogged better.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

FIG1 is a flow chart of an image defogging method based on a combination of prior knowledge and deep learning according to an embodiment of the present invention.

DETAILED DESCRIPTION

The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

Referring to FIG. 1 , the present invention describes a single image defogging method based on a priori knowledge combined with deep learning, comprising the following steps:

Decompose the input original foggy image Z using a weighted guided filter to obtain a detail layer image _Ze and a base layer image _Zb ;

Processing the base layer image Z _b using a quadtree search method to obtain a global atmospheric light component image A _c ;

Constructing a deep convolutional neural network to process the base layer image Z _b to obtain a transmittance image t;

Based on the atmospheric scattering model, the original foggy image Z is restored using the global atmospheric light component image A _c , the transmittance image t and the detail layer image _Ze to obtain a defogged image J.

Specifically, the method of decomposing the input original foggy image Z by using a weighted guided filter to obtain a detail layer image _Ze and a base layer image _Zb specifically includes the following steps:

Get the pixel gray value Z(x,y) at any point (x,y) in the original foggy image Z;

The weighted filter coefficients a _p (x, y) and _bp (x, y) at the midpoint (x, y) of the original foggy image Z are obtained using the following formula:

Wherein, μ _Z,ρ (x, y) is the mean value of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z;

是在原始有雾图像Z中以点(x，y)为中心，ρ为半径的窗体的像素灰度值的方差；

It is the variance of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z;

Γ _Y (x, y) is the brightness component of the point (x, y) of the original foggy image Z;

λ is a constant greater than 1;

The base layer image Z _b is obtained using the following formula:

Z _b (x, y) = a _p (x, y) Z (x, y) + b _p (x, y);

Wherein, Z _b (x, y) is the grayscale value of the pixel at the point (x, y) in the base layer image Z _b ;

Z(x,y) is the grayscale value of the pixel at any point (x,y) in the original foggy image Z;

a _p (x, y) and _bp (x, y) are the weighted filter coefficients at the midpoint (x, y) of the original foggy image Z.

Specifically, the method of using the quadtree search method to process the base layer image Z _b to obtain the global atmospheric light component image A _c specifically includes the following steps:

Step A, divide the base layer image Z _b into four rectangular regions Z _bi (i∈{1,2,3,4}), the length and width of Z _bi are 1/2 of the length and width of Z _b respectively;

Step B, defining the score of each rectangular area as the mean value of the pixel grayscale values minus the standard deviation of the pixel grayscale values in the area;

Step C, selecting the rectangular area with the highest score and further dividing it into four rectangular areas;

Repeat steps B and C and iteratively update the rectangular area with the highest score n times to obtain the final subdivided area Z _b-end ;

The atmospheric light component image A _c is obtained using the following formula:

|A _c (c∈{r,g,b})|=min|((Z _b-end-r (x,y),Z _b-end-g (x,y),Z _b-end-b (x,y))-(255,255,255))|,

Wherein (Z _b-end-r (x, y), Z _b-end-g (x, y), Z _b-end-b (x, y)) is the color vector of any pixel point of the final subdivision area Z _b-end , and (255, 255, 255) is a pure white vector.

Specifically, the construction of a deep convolutional neural network to process the base layer image _Zb to obtain the transmittance image t specifically includes the following steps:

The base layer image Z _b is downsampled by using the first convolutional layer to obtain a low-resolution base layer image Z _b ' and then the low-level features are extracted, wherein the formula for downsampling the base layer image Z _b by using the first convolutional layer is:

为第i层卷积层在通道c的索引下，所述低分辨率基础层图像Z_b’的低层特征；in,

is the low-level features of the low-resolution base layer image Z _b ′ at the index of channel c at the i-th convolutional layer;

x is the horizontal coordinate of the base layer image Z _b ;

y is the ordinate of the base layer image Z _b ;

x' is the abscissa of the low-resolution base layer image Z _b ';

y' is the ordinate of the low-resolution base layer image Z _b ';

为第i层卷积层在c和c'图层索引通道下的卷积权重数组；

is the convolution weight array of the i-th convolution layer under the c and c' layer index channels;

为第i层卷积层在c图层索引通道下的偏差向量；

is the bias vector of the i-th convolutional layer under the c-th layer index channel;

σ(·)＝max(·,0) represents the ReLU activation function, and zero padding is used as the boundary condition of all convolutional layers;

s is the step size of the convolution kernel;

输入层数为n_L＝2的第二卷积层提取局部特征L，其中第二卷积层中卷积核的尺寸为3×3，步长为1；The obtained low-level features

The second convolutional layer with the input layer number n _L = 2 extracts the local feature L, where the size of the convolution kernel in the second convolutional layer is 3×3 and the stride is 1;

输入层数为n_G1＝2、卷积核的尺寸为3×3和步长为2的第三卷积层后，再输入层数为n_G2＝2的全连接层，得到全局特征G；The low-level features obtained by the second convolutional layer

After inputting the third convolutional layer with n _G1 = 2 layers, a convolution kernel size of 3×3 and a stride of 2, the fully connected layer with n _G2 = 2 layers is input to obtain the global feature G;

对应的混合特征图F_L＝σ(L+G)；Add the local feature L and the global feature G and input them into the activation function to get the low-level feature

The corresponding mixed feature map F _L =σ(L+G);

对应的初步大气折射率特征图，然后再对所述初步大气折射率特征图进行上采样，输出得到透射率图像t(x,y)。Convolution is performed on the mixed feature map F _L =σ(L+G) to obtain the low-level features

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to output a transmittance image t(x, y).

Specifically, based on the atmospheric scattering model, the global atmospheric light component image A _c , the transmittance image t and the detail layer image _Ze are used to restore the original foggy image Z, and the formula for obtaining the defogging image J is:

Where J(x,y) is the grayscale value of the pixel at the midpoint (x, y) of the dehazed image J.

_Ze (x,y) is the grayscale value of the pixel at the midpoint (x, y) of the detail layer image _Ze .

Z _b (x, y) is the grayscale value of the pixel at the midpoint (x, y) of the base layer image Z _b ,

t(x,y) is the output transmittance image,

其中t(x,y)为所述输出得到透射率图像，η为大于零的常数。

Wherein t(x, y) is the output transmittance image, and η is a constant greater than zero.

In this step, the foggy image is restored according to the atmospheric scattering model Z(x,y)=t(x,y)J(x,y)+A _c (1-t(x,y)), where J(x,y) is the defogging image and (x,y) is the spatial coordinate of the pixel point. The restored image can be expressed as:

噪声将被放大，考虑噪声主要包含在细节层图像Z_e中以及噪声对图像影响，本发明恢复后发的图像表示为，Among them, t ₀ is a parameter to ensure the processing effect of dense fog area, Z(x,y)＝J(x,y)+n(x,y), J is the noise-free image, n is the noise, that is,

The noise will be amplified. Considering that the noise is mainly contained in the detail layer image _Ze and the influence of the noise on the image, the image after restoration of the present invention is expressed as:

是为了降低噪声对恢复图像的影响，其表示为，The atmospheric light component A is obtained through step 2.1, and the transmittance t is obtained through step 2.2.

It is to reduce the impact of noise on the restored image, which is expressed as,

避免天空区域的噪声被放大；η is a constant. In this embodiment, η＝6. Experiments show that t(x,y)∈[0,1], when t(x,y)＜1/η, the pixel point in the sky area of (x,y) data is

Avoid amplifying noise in the sky area;

对应的初步大气折射率特征图，然后再对所述大气折射率特征图进行上采样，输出得到透射率图像t之后还包括如下步骤：Specifically, the mixed feature map F _L =σ(L+G) is convolved to obtain the low-level feature

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to obtain the transmittance image t, and then the following steps are included:

The loss function is constructed as follows:

L＝ _{Lr + wcLc}

L_c表示为

Among them, _Lr is the reconstruction loss function, _Lc is the color loss function, _wc is the weight assigned to the color loss function _Lc , and _Lr is expressed as

L _c is expressed as

J is the defogged image, Z is the foggy image, c∈(R,G,B) is the channel index, ∠(J(x,y),Z(x,y)) represents the angle between the three-dimensional color vectors of the foggy image and the defogged image at the pixel point (x,y). Although the reconstruction error can measure the similarity between the original image and the defogged image, it cannot ensure that the angles of their color vectors are consistent. Therefore, a color error function is added to ensure that the angles of the color vectors of the same pixel point in the image before and after defogging are consistent.

The beneficial effect of the image defogging method and device based on the combination of prior knowledge and deep learning provided by the present invention is that the foggy image is decomposed by a weighted guided filter to obtain a detail layer image and a base layer image, and then the base layer image is processed by a quadtree search method and a deep neural network to obtain a global atmospheric light component image and a transmittance image, and finally the foggy image is restored by using the global atmospheric light component image, the transmittance image and the detail layer image to obtain a defogged image. Since both the global atmospheric light component image and the transmittance image use the base layer image for estimation or calculation, the amplification of image noise is avoided, and the foggy image can be defogged better. The loss function is calculated using the defogged image for feedback parameter adjustment, and color loss is added to the loss function to improve the robustness of the algorithm.

In another embodiment of the present invention, an image defogging device based on a priori knowledge combined with deep learning includes a memory and a processor. The memory is used to store a computer program. The processor is used to implement the image defogging method based on a priori knowledge combined with deep learning as described above when executing the computer program.

The reader should understand that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. Those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples. Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be construed as limitations on the present invention, and those of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims (6)

1. An image defogging method based on a priori knowledge combined with deep learning, characterized in that the method comprises: Decompose the input original foggy image Z using a weighted guided filter to obtain a detail layer image _Ze and a base layer image _Zb ; The base layer image Z _b is processed by a quadtree search method to obtain a global atmospheric light component image A _c , where c∈{r,g,b}; Constructing a deep convolutional neural network to process the base layer image Z _b to obtain a transmittance image t; Based on the atmospheric scattering model, the original foggy image Z is restored using the global atmospheric light component image A _c , the transmittance image t and the detail layer image _Ze to obtain a defogged image J; The formula for restoring the original foggy image Z based on the atmospheric scattering model using the global atmospheric light component image _Ac , the transmittance image t and the detail layer image _Ze to obtain the defogging image J is:

Wherein, J(x,y) is the pixel grayscale value at the midpoint (x, y) of the dehazed image J, _Ze (x,y) is the pixel grayscale value at the midpoint (x, y) of the detail layer image _Ze , _Zb (x,y) is the pixel grayscale value at the midpoint (x, y) of the base layer image _Zb , t(x,y) is the transmittance at the midpoint (x, y) of the transmittance image t,

Where η is a constant greater than zero. 2. The image defogging method based on the combination of prior knowledge and deep learning according to claim 1 is characterized in that the input original foggy image Z is decomposed by using a weighted guided filter to obtain a base layer image Z _b , which specifically comprises the following steps: Get the pixel gray value Z(x,y) at any point (x,y) in the original foggy image Z; The weighted filter coefficients a _p (x, y) and _bp (x, y) at the midpoint (x, y) of the original foggy image Z are obtained using the following formula:

Wherein, μ _Z,ρ (x, y) is the mean value of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z;

is the variance of the pixel grayscale value of the window with point (x, y) as the center and ρ as the radius in the original foggy image Z; Γ _Y (x, y) is the brightness component of the original foggy image Z at point (x, y), Y = max{Z _r , Z _g , Z _b }, Z _r , Z _g , Z _b are the R, G, B values at point (x, y) in the foggy image Z respectively; λ is a constant greater than 1; The base layer image Z _b is obtained using the following formula: Z _b (x, y) = a _p (x, y) Z (x, y) + b _p (x, y); Wherein, Z _b (x, y) is the grayscale value of the pixel at the point (x, y) in the base layer image Z _b . 3. The image defogging method based on the combination of prior knowledge and deep learning according to claim 1 is characterized in that the base layer image Z _b is processed by a quadtree search method to obtain a global atmospheric light component image A _c , which specifically comprises the following steps: Step A, divide the base layer image Z _b into four rectangular regions Z _bi (i∈{1,2,3,4}), the length and width of Z _bi are 1/2 of the length and width of Z _b respectively; Step B, defining the score of each rectangular area as the mean value of the pixel grayscale values minus the standard deviation of the pixel grayscale values in the area; Step C, selecting the rectangular area with the highest score and further dividing it into four rectangular areas; Step D, repeating steps B and C and iteratively updating the rectangular area with the highest score n times to obtain the final subdivided area Z _b-end ; The atmospheric light component image A _c is obtained using the following formula: |A _c (c∈{r,g,b})|=min|((Z _b-end-r (x,y),Z _b-end-g (x,y),Z _b-end-b (x,y))-(255,255,255))|, Wherein (Z _b-end-r (x, y), Z _b-end-g (x, y), Z _b-end-b (x, y)) is the color vector at the midpoint (x, y) of the final subdivision area Z _b-end , and (255, 255, 255) is a pure white vector. 4. The image defogging method based on the combination of prior knowledge and deep learning according to claim 1 is characterized in that the construction of a deep convolutional neural network processes the base layer image Z _b to obtain the transmittance image t, specifically comprising the following steps: The base layer image Z _b is downsampled by using the first convolutional layer to obtain a low-resolution base layer image Z _b ' and then extract low-level features, wherein the formula for downsampling the base layer image Z _b by using the first convolutional layer is:

in,

is the low-level feature of the low-resolution base layer image Z _b ′ under the index of channel c of the i-th convolution layer in the first convolution layer; x is the abscissa of the base layer image Z _b , and y is the ordinate of the base layer image Z _b ; x' is the abscissa of the low-resolution base layer image Z _b ′, and y' is the ordinate of the low-resolution base layer image Z _b ′;

is the convolution weight array of the i-th convolution layer in the first convolution layer under the c and c' layer index channels;

is the bias vector of the i-th convolutional layer in the first convolutional layer under the c-th layer index channel; σ(·)＝max(·,0) represents the ReLU activation function, and zero padding is used as the boundary condition of all convolutional layers in the first convolutional layer; s is the step size of the convolution kernel of the first convolutional layer; The obtained low-level features

The second convolutional layer with the input layer number n _L = 2 extracts the local feature L, where the size of the convolution kernel in the second convolutional layer is 3×3 and the stride is 1; The low-level features obtained by the second convolutional layer

After inputting the third convolutional layer with n _G1 = 2 layers, a convolution kernel size of 3×3 and a stride of 2, the fully connected layer with n _G2 = 2 layers is input to obtain the global feature G; Add the local feature L and the global feature G and input them into the activation function to get the low-level feature

The corresponding mixed feature map F _L =σ(L+G); Convolution is performed on the mixed feature map F _L =σ(L+G) to obtain the low-level features

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to output a transmittance image t. 5. The image defogging method based on prior knowledge and deep learning according to claim 4 is characterized in that the mixed feature map F _L =σ(L+G) is convolved to obtain the low-level feature map

The corresponding preliminary atmospheric refractive index characteristic map is then upsampled to obtain the transmittance image t, and then the following steps are included: The loss function is constructed as follows: L＝ _{Lr + wcLc} Among them, _Lr is the reconstruction loss function, _Lc is the color loss function, _wc is the weight assigned to the color loss function _Lc , and _Lr is expressed as

L _c is expressed as

J is the defogging image, Z is the foggy image, c∈(R,G,B) is the channel index, ∠(J(x,y),Z(x,y)) represents the angle between the three-dimensional color vectors of the foggy image and the defogging image at the pixel point (x,y); The loss function is used to adjust the parameters of the deep convolutional neural network. 6. An image defogging device based on a priori knowledge combined with deep learning, characterized in that the device includes a memory and a processor; The memory is used to store computer programs; The processor is used to implement the image defogging method based on the combination of prior knowledge and deep learning as described in any one of claims 1 to 5 when executing the computer program.

Images