CN107798665B - Underwater image enhancement method based on structure-texture layering - Google Patents

Abstract

The invention belongs to the field of computer vision, and aims to improve the contrast of an image and keep good details, inhibit noise and keep the naturalness of color tones. The invention adopts the technical scheme that an underwater image enhancement method based on structure-texture layering is characterized in that firstly, color correction is carried out through histogram equalization, and then a color corrected image is decomposed into a low-frequency structure layer and a high-frequency texture layer. The method comprises the steps of enabling noise to be remained in a texture layer, accurately estimating the transmittance from a structure layer without noise based on a proposed fog line model, further performing enhancement processing, enhancing the texture layer by using a gradient residue minimization method, and reconstructing the enhanced structure layer, the texture layer and a refined edge mask into a final enhancement image through proper scale expansion, so that the problem which cannot be processed in the prior art is solved. The invention is mainly applied to underwater image enhancement and processing occasions.

Description

Underwater image enhancement method based on structure-texture layering

technical field

The invention belongs to the field of computer vision, and relates to an underwater image enhancement method based on structure-texture layering. Specifically, the image is divided into a high-frequency texture layer and a low-frequency structure layer by layering, which not only avoids the problem of amplifying noise during the enhancement process, but also preserves details and edges, improving contrast and tone naturalness. , which is an underwater image enhancement method based on structure-texture layering.

Background technique

Clear and high-visibility underwater image analysis and recognition is of great significance to research in underwater surveying, underwater video, and military research. However, in the complex underwater imaging environment, the image quality is severely degraded due to the effect of absorption and scattering by large suspended particles (like turbid particles in water) during transmission. Therefore, typical underwater images have the following problems:

1) Concentrated light illumination, resulting in uneven distribution of background grayscale;

2) Absorption and scattering effects make underwater images non-uniform brightness and detail blurring;

3) The lighting conditions are poor, causing the problem of tone attenuation and contrast reduction;

Since the quality of the underwater images captured is generally poor, preprocessing should be performed before target recognition is performed on the underwater images. The existing underwater image enhancement methods mainly have the following development stages:

The first stage: use traditional image enhancement methods to process underwater images, such as: grayscale changes, histogram equalization, image spatial domain smoothing and sharpening, and pseudo-color processing. However, due to the unique characteristics of underwater imaging, the results of these traditional methods are not ideal.

The second stage: According to the similarity of imaging characteristics in underwater and haze, researchers began to study underwater image enhancement methods based on dehazing technology. However, due to the difference in transmittance and absorption and scattering characteristics in water and air, There are some problems in directly applying the dehazing method to underwater image enhancement, such as loss of details in dark areas and overexposure in bright areas.

The third stage: With the study of underwater refraction and scattering properties, researchers have proposed some specialized underwater mathematical imaging models and some methods for underwater image enhancement. The main starting points of these methods are background light estimation, transmittance estimation, compensation for light wave attenuation, enhancement based on fusion principle, etc. For example: the background light is estimated by the layer-based method mentioned in the reference; the estimation method of the global background light based on quadtree decomposition and image segmentation, and a method for the minimum information loss of underwater imaging characteristics is proposed. A transmittance estimation algorithm; a systematic approach to enhancing underwater images through artificial light compensation methods; increasing the visibility of underwater degraded images by fusing four weights. These methods suppress the overexposure and noise problems to a certain extent, and also improve the color tone of the image, but there are still some problems such as loss of details, noise amplification, unnatural color tone, and oversaturation. These problems will have a great impact on underwater research. Therefore, a better underwater image enhancement method is designed to improve the image contrast as much as possible, highlight the details and features of interest, improve the visual effect of the image, provide clarity , An image suitable for analysis is necessary.

SUMMARY OF THE INVENTION

The present invention is intended to make up for the deficiencies of the prior art, that is, to improve the contrast of the image while retaining good details, and to suppress the noise while maintaining the naturalness of the color tone. The technical scheme adopted in the present invention is that, for an underwater image enhancement method based on structure-texture layering, first, color correction is performed through histogram equalization, and then the color-corrected image is decomposed into a low-frequency structure layer and a high-frequency texture layer . Residual noise to the texture layer, and then based on the proposed fog line model, accurately estimated transmittance from the structure layer without noise, and then perform enhancement processing, and then enhance the texture layer with the gradient residual minimization method. The scale scaling of the enhanced structure layer, texture layer and refined edge mask is reconstructed into the final enhanced map, thus solving the problem that the existing technology cannot handle.

Specific steps are as follows:

按照以下方法计算边界的下限

和上限

1) Perform color correction on the captured image: First, combined with the characteristics of the color distribution histogram of the underwater image, move the average color distribution of each color channel to the desired range, and then perform linear normalization, that is, use a simple The histogram equalization of the

Calculate the lower bound of the boundary as follows

and cap

where ^c∈ {R,G,B}, μc and ^σc are the mean and standard deviation of each color channel, respectively, λ is a hue parameter that effectively adjusts the distribution range, and then uses the truncation function chip() to convert The pixel value range of the image is truncated to [0, 255], resulting in a color-corrected image:

I ^c (x) is the pixel value of each channel of the color-corrected image;

2) Build a layered model: Define an underwater image with a dehazing model:

I ^c (x)=J ^c (x)t(x)+A ^c (1-t(x))+E(x), (4)

Where, x is the pixel value, J ^c (x) is the restored image, A ^c is the global background light, t(x) is the projection rate in water, E(x) is defined as the noise in the input image, and the image is enhanced The purpose is to recover J ^c (x) from the captured underwater image I ^c (x), which involves the estimation of transmittance and background light;

3) Structure-texture layering implementation: TV-L ¹ is used to separate the structure layer and the texture layer:

是通过1范数正规化结构层的梯度，ξ是控制光滑程度的权重参数，通过TV-L¹将结构层I_s求解出来，TV-L¹是将全变分作为正则项进行图像分解的最优化方法，纹理层通过I_t＝I-I_s算出；in,

is the gradient of the structural layer normalized by the 1-norm, ξ is the weight parameter controlling the smoothness, the structural layer I _s is solved by TV-L ¹ , and TV-L ¹ uses the total variation as a regular term to decompose the image Optimization method, the texture layer is calculated by It = II _{s ;}

4) Structural layer enhancement: Use the global background light scattering estimation method to estimate A:

是达到结构层的暗通道中最大值位置，即

Ω(y)是y的一个领域；in,

is the maximum position in the dark channel that reaches the structural layer, that is,

Ω(y) is a field of y;

According to the dark channel prior theory, first assume that the transmittance in each window is constant, and define it as t ₀ . The value of A has been calculated by formula (9), so the minimum value is calculated twice on both sides of formula (6). , the estimated value of transmittance t(x) is obtained after shifting the items:

where ω is the correction factor;

5) Texture layer denoising: Through the above TV-L ¹ decomposition, the noise in the captured image remains in the high-frequency texture layer. Gradient residual minimization is used to enhance the texture layer. The enhancement of this texture layer is obtained by optimizing formula (15):

Among them, δ and η are the control weight parameters, which control the smoothness and fineness of the processing result. If η is too large or δ is too small, most of the texture of the enhanced image will be lost with the denoising process. Therefore, it is appropriate to select δ The value of and η has a great influence on the experimental results. Let Z=J-I, then the above optimization problem becomes two sub-problems, as follows:

This turns Equation (16) into a classic TV problem, and Equation (17) is approximated with a soft threshold;

6) Mask refinement: Use a binary mask M to separate the texture layer decomposed by TV-L ¹ into smooth areas and detail areas, and then delete the residual details in the smooth area of the enhanced low-frequency structure layer. Cosine transform coefficient to detect the similarity between blocks in the scene, to judge whether a region is smooth, define the _DCT coefficient of each 8 × 8 block in It as B, then, the similarity of each block in the scene details Sex is expressed by formula (18):

采用软映射的方法结合结构层I_s产生的映射拉布拉氏矩阵修正这初始的粗糙的M：Among them, x, y are the coordinate positions in the DCT, calculate the square sum of all DCT coefficients except B _1,1 , B _1,2 and B _2,1 , and then use the threshold to judge the similarity of each block, define The initial edge mask of the texture layer is M, and the threshold κ is set to 0.1. When ρ is greater than κ, M of the block is 1, otherwise, M is 0, so that a rough binary edge information map is obtained. In order to obtain finer edge mask

The initial rough M is modified by the soft mapping method combined with the mapping _Labrador matrix generated by the structural layer Is:

是正规化参数，在文中设为10⁵；where m and m' are the vector representations of M and M', L _s is the mapped _Labrador matrix generated from Is,

is the normalization parameter, which is set to 10 ⁵ in the text;

7) Reconstruction: Finally, we reconstruct the final enhancement processing result by formula (20):

和

分别是经过增强处理的低频结构层和高频纹理层。where τ is the scaling factor, which introduces enhanced details, and is set as

and

They are the enhanced low-frequency structure layer and high-frequency texture layer, respectively.

In the low-light environment, the high-frequency part of the image not only contains texture information, but also contains a lot of noise. In order to avoid affecting the accuracy of transmittance estimation, the image is first processed in layers, and then the transmittance is estimated at the low-frequency structure layer. The high-frequency texture layer is denoised, and a picture is defined as J=J _s +J _t , where J _s and J _t are the structure layer and texture layer of the image, respectively, then equation (4) is rewritten as:

I ^c (x)=(J _s ^c (x)+J _t ^c (x))t(x)+A ^c (1-t(x))+E(x)

=J _s ^c (x)t(x)+A ^c (1-t(x))+J _t ^c (x)t(x)+E(x).(5)

The background light ^Ac is a relatively smooth signal, and the estimated transmittance t is also close to the ideal transmittance in J _s ^c . Therefore, the captured image can be written as: I=I _s +I _t , and we get:

I _s ^c (x)=J _s ^c (x)t(x)+A ^c (1-t(x)) (6)

I _t ^c (x)=J _t ^c (x)t(x)+E(x) (7)

Among them, I _{sc is the low-frequency structure layer of the captured image, which contains most of the structure contours of the image, and I t c} ^is the high-frequency texture layer of the captured image, which contains most of the texture information and noise ^. Through this method, the transmittance _t (x) is estimated from Is (x) and is immune to noise.

The estimated value of transmittance t(x) is more refined. The processing steps are: transform formula (6) into the 3D coordinate system, and obtain a linear transmittance estimate value t'(x) of transmittance in the three-dimensional coordinate system, Taking the background light A as the origin of the coordinates, formula (6) is transformed into the following form:

I _s0 (x)=t'(x) · J _s0 (x) (11)

where I _s0 (x)=I _s (x)-A, J _s0 (x)=J _s (x)-A, therefore, the linear transmittance estimate for each pixel value is calculated according to the following formula:

紧接着，通过优化公式(13)，进而得到一个精细的透射率映射：Given a boundary value of transmittance

Next, by optimizing formula (13), a fine transmittance map is obtained:

在每一个像素值上的标准偏差，N(x)表示在结构层中每一个像素值周围的四邻域像素位置，A和

的值被计算出来之后，那么，这个增强的结构层可以由公式(14)得出：where α and β are the weight parameters that control the data item and the smooth item, and σ(x) is

The standard deviation at each pixel value, N(x) represents the four-neighborhood pixel location around each pixel value in the structure layer, A and

After the value of is calculated, then, this enhanced structural layer can be obtained by Equation (14):

Technical features and effects of the present invention:

The method of the invention aims at the problems of noise amplification and tone shift of underwater image enhancement. By dividing the image into a structural layer and a texture layer, and performing enhancement and denoising respectively, the contrast of the image is improved while good details are preserved, and the noise is suppressed. While maintaining the naturalness of the tones. The present invention has the following characteristics:

1. The main innovation point of the algorithm proposed in this paper is based on the idea of structure-texture layering, and the two layers are enhanced and denoised respectively to avoid the influence of noise amplification.

2. This paper proposes a simple and universally applicable solution to the color shift problem.

3. It has good results for various underwater graphics enhancement, and has certain universality.

4. The algorithm proposed in this paper can be applied to the enhancement of underwater video, and has certain scalability.

Description of drawings

Figure 1 is the algorithm framework diagram;

The left picture of Figure 2 is the captured underwater image, and the right picture is the corresponding histogram;

The left picture of Figure 3 is the image after color correction, and the right picture is the corresponding histogram;

Figure 4 is the decomposition model verification diagram: Figure 4 ((a) synthesized degraded image I; (b) is the structural layer of I; (c) is the texture layer of I; (d) is a clear image J; (e) ) is the structural layer of J; (f) is the texture layer of J);

Fig. 5 is the decomposition result diagram;

Fig. 6 is the enhancement result diagram of the structural layer;

Fig. 7 is the enhancement result diagram of texture layer;

Figure 8 is a redefinition diagram of an edge mask;

Fig. 9 is the result picture of final reconstruction;

Figure 10 is a comparison diagram of the effects of this article and other references (the first column is the input image; the second, third and fourth columns are the results of three examples of prior art; the fifth column is the results of this article);

Figure 11 is a layered and non-layered enhancement effect map (the first row is the result map without layering; the second row is the result map corresponding to layering);

Detailed ways

The present invention proposes an underwater image enhancement method based on structure-texture layering, and the noise remains in the texture layer, and the structure layer without the influence of noise satisfies the fog line model. Then we enhance the texture layer by using the fog-line based model and enhance the texture layer with gradient residual minimization method. Afterwards, the enhanced structure layer, texture layer and refined mask are reconstructed into the final enhanced map through appropriate scaling. That is, to improve the contrast of the image while retaining good details, and to suppress the noise while maintaining the naturalness of the tone. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

First, in view of some shortcomings of the existing underwater image enhancement methods, we propose a layer-based image processing method, the frame of which is shown in Figure 1. The first step is color correction through histogram equalization. The second step is to use TV-L ¹ (the optimization method of image decomposition (called ROF or TV-L ¹ model) with total variation (TV energy) as a regular term) method to stratify the image, and then to The structure layer and texture layer are respectively enhanced and denoised. Finally, the enhanced structural and texture layers and the refined mask are reconstructed into the final result map. The following is the specific implementation method:

1) Perform color correction on the captured image. By observing the characteristics of the color distribution histogram of the underwater image as shown in Figure 2,

(

代表图像每个色彩通道的像素值)，按照以下方法计算边界的下限

(

代表图像每个色彩通道的像素值的最小值)和上限

(

代表图像每个色彩通道的像素值的最大值)：8) We see that because in water, different light segments suffer from different intensities of absorption and scattering, only the blue-green light in the 480-570 band has the smallest attenuation coefficient and the strongest penetration ability in water, so more images are captured in water. Appears in a light green light tint. To address this color shift problem, we employ simple histogram equalization. The principle is very intuitive, that is, move the average color distribution of each color channel to the desired range, and then perform linear normalization. for each color channel

(

represents the pixel value of each color channel of the image), calculate the lower bound of the boundary as follows

(

Represents the minimum value) and upper limit of the pixel value of each color channel of the image

(

represents the maximum value of the pixel values for each color channel of the image):

where ^c∈ {R,G,B}, μc and ^σc are the mean and standard deviation of each color channel, respectively. λ is a hue parameter that effectively adjusts the distribution range beyond which pixel values are truncated. Finally, this color-corrected image is obtained by minimizing equation (3).

I ^c (x)=J ^c (x)t(x)+A ^c (1-t(x))+E(x), (4)

where clip( ) is a clip function that controls the truncation of the pixel value range to [0, 255]. Through the comparison of Figure 2 and Figure 3, it can be shown that this method can effectively correct the color of underwater images.

2) Build a hierarchical model. Since the attenuation of the optical path in turbid water is similar to the attenuation in the atmosphere with haze, He Kaiming's classic dark channel dehazing model is used to define the underwater graph model:

I ^c (x)=J ^c (x)t(x)+A ^c (1-t(x))+E(x), (4)

where x is the pixel value, ^Ic (x) is the color-corrected image for each channel, and ^Jc (x) is the restored image. A ^c is the global background light, and t(x) is the transmittance in the water. Define E(x) as the noise in the input image. The purpose of image enhancement is to recover J ^c (x) from the captured underwater image I ^c (x), which involves the estimation of transmittance and background light

Usually, images not only contain important texture information in high-frequency parts, but also contain a lot of noise, especially in low-light environments. These factors reduce the accuracy of the transmittance estimation, which in turn affects the final enhancement effect. In order to overcome this problem, the image is first processed in layers, and then the transmittance is estimated in the structure layer to avoid the influence of noise, and denoising is performed separately in the texture layer. We define J _s and J _t as the structure layer and texture layer of the clean image, respectively, and then have J = J _s + J _t , then equation (4) can be rewritten as:

I ^c (x)=(J _s ^c (x)+J _t ^c (x))t(x)+A ^c (1-t(x))+E(x)

=J _s ^c (x)t(x)+A ^c (1-t(x))+J _t ^c (x)t(x)+E(x).(5)

The background light ^Ac is a smooth signal, and the estimated transmittance _t is also relatively smooth, close to the ideal transmittance in J ^sc . Therefore, the observed captured image can be divided into two layers (I _{= Is} +It), which ideally can be obtained as:

I _s ^c (x)=J _s ^c (x)t(x)+A ^c (1-t(x)) (6)

Among them, I _sc is the structural ^layer of the captured image, which contains most of the details of the image. I _t ^c is the texture layer of the captured image, which contains most of the texture information and noise. By this method, the transmittance _t (x) can be estimated from Is (x), free from the influence of noise.

3) Implement layering. Because the structure layer has a larger gradient in the contour and boundary of the object, but has a smaller gradient in the texture layer. We use the TV-L ¹ decomposition model to nicely separate the structural and texture layers:

是通过1范数正规化结构层的梯度，ξ是控制光滑程度的权重参数，通过TV-L¹(将全变分(TV能量)作为正则项进行图像分解(称为ROF或TV-L¹模型)的最优化方法)可以将结构层I_s求解出来，纹理层通过I_t＝I-I_s算出。in,

is the gradient of the structured layer normalized by the 1-norm, ξ is the weight parameter that controls the degree of smoothness, and the image decomposition (called ROF or TV-L ¹ ) is performed by TV-L ¹ (with the total variation (TV energy) as the regular term). The optimization method of the model) can solve the structure layer Is _, and the texture layer can _{be calculated by It = II s .}

An important question is whether the above TV ^- L1 decomposition model can exactly decompose the captured graphs according to Equations (6) and (7). Figure 4 shows an example of decomposition, this input image is I synthesized from J according to equation (4), J _s and J _t are the two-layer images decomposed with TV-L ¹ optimization, as shown in the figure, The structural layer (I _s ) of the degraded image I(x) excludes high-frequency texture details and noise, and is visually consistent with the structural layer of the original clear image. In order to have better visual effect, we enlarge the texture layer by 10 times.

4) Background light estimation. Taking into account the absorption and scattering effects of light waves in water, we estimate A using the global background light scattering estimation method:

是达到结构层的暗通道中最大值位置，即

Ω(y)是y的一个领域，它的尺寸设为17。in,

is the maximum position in the dark channel that reaches the structural layer, that is,

Ω(y) is a field of y whose size is set to 17.

According to the dark channel prior theory, first assume that the transmittance in each window is constant, define it as t ₀ , and the A value has been calculated by formula (9), so we find the minimum value twice on both sides of formula (6) After the operation, the estimated value of the transmittance t(x) can be obtained:

Among them, ω is the correction factor, which is set to 0.95 in this paper, the purpose is to make the restoration result map closer to the realistic photo. However, there are obvious inconsistencies in the result image recovered by only using this estimated value of transmittance t(x), so in order to obtain a better enhancement effect, we need to refine this estimated value. We _observe that the distribution of pixels in the Is of the distribution without noise conforms to the "fog line" law. Therefore, we can convert the formula (6) into the 3D coordinate system to obtain a transmittance t'(x) in the three-dimensional coordinate system (hereinafter referred to as the linear transmittance estimation value). Taking the background light A as the coordinate origin, the formula (6) can be transformed into the following form:

I _s0 (x)=t'(x) · J _s0 (x) (11)

Wherein, I _s0 (x)=I _s (x)-A, and J _s0 (x)=J _s (x)-A. Therefore, we can calculate the linear transmittance estimate for each pixel value according to the following formula:

紧接着，我们可以通过优化公式(12)，进而得到一个精细的透射率映射：Given a boundary value of transmittance

Next, we can obtain a fine transmittance map by optimizing formula (12):

在每一个像素值上的标准偏差，N(x)表示在结构层中每一个像素值的四领域像素位置。一旦A和

的值被计算出来，那么，这个增强的结构层可以由公式(14)得出：where α and β are the weight parameters that control the data item and the smooth item, and σ(x) is

The standard deviation at each pixel value, N(x) represents the four-domain pixel location for each pixel value in the structured layer. Once A and

The value of is calculated, then, this enhanced structural layer can be derived from equation (14):

5) Texture layer enhancement. After the above TV-L ¹ decomposition, the noise in the captured image remains in the texture layer. We use the gradient residual minimization method to enhance the texture layer, that is, the enhanced texture layer can be obtained by optimizing the formula (15):

Among them, δ and η are the control weight parameters, which control the smoothness and fineness of the processing results. If η is too large or δ is too small, most of the texture of the enhanced image will be lost with the denoising process. Therefore, it is appropriate to select δ and δ. The value of η has a great influence on the experimental results. Let Z=J-I, the above optimization problem can be transformed into two sub-problems, as follows.

Equation (16) becomes a classic TV problem, which can be solved by referring to the TV-L ¹ method in [15]. Equation (17) can be approximated with a soft threshold.

可能还包括一部分不希望存在的噪声，尤其在光滑区域更加明显。因此，我们用一个边缘掩膜M把TV-L¹分解后的纹理层分离成光滑和细节区域，进而删除增强后结构层中光滑区域的残留细节。我们用离散余弦变换系数来检测场景中块之间的相似性，判断是否光滑，定义I_t中的每一个8×8的块的DCT系数为B,然后，场景细节中的每个块的相似性可以用公式(18)表示：6) Definition of edge mask. It can be seen from Figure 5 that the enhanced

It may also include some unwanted noise, especially in smooth areas. Therefore, we use an edge mask M to separate the texture layer after TV-L ¹ decomposition into smooth and detail regions, and then remove the residual details in the smooth region in the enhanced texture layer. We use discrete cosine transform coefficients to detect the similarity between blocks in the scene, judge whether it is smooth, define the _DCT coefficient of each 8 × 8 block in It as B, and then, the similarity of each block in the scene details Sex can be expressed by formula (18):

我们软映射的方法结合结构层I_s产生的映射拉布拉氏矩阵修正这初始的粗糙的M。Among them, x, y are the coordinate positions in the DCT, we calculate the square sum of all DCT coefficients except B _1,1 , B _1,2 and B _2,1 , and then use the threshold to judge the similarity of each block . We define the initial detail mask of the texture layer as M, the threshold κ is 0.1, when ρ is greater than κ, the block's M is 1, otherwise, M is 0. To get fine detail mask

Our soft-mapping method modifies this initial rough M by combining the mapping _Labrador matrix produced by the structural layer Is.

是正规化参数，在文中设为10⁵。where m and m' are the vector representations of M and M', L _s is the mapped _Labrador matrix generated from Is,

is the normalization parameter, which is set to 10 ⁵ in the text.

7) Reconstruct the image. The structure layer and the texture layer are strengthened respectively above, and the edge mask M is obtained, so the final enhancement processing result can be reconstructed by formula (20):

和

分别是被增强的结构层和纹理层。where τ is the scaling factor, which introduces enhanced details, and is set as

and

They are the enhanced structural layer and texture layer, respectively.

In the experiment, the color correction effect, layering effect, denoising effect and enhancement effect of the algorithm in this paper were tested, and compared with the algorithm of the comparative literature. The specific results are shown in Figure 2, Figure 3, Figure 4, Figure 5, and Figure 5. 6. As shown in Figure 7, Figure 9, Figure 10, and Figure 11. It can be seen in FIG. 10 that the restoration effect of the prior art example 1 amplifies the noise, and the tone of the enhanced image is unnatural. The restoration result of the prior art example 2 has a relatively low contrast and obvious noise. The processing result of the prior art example 3 improves the contrast of the image and reduces the noise to a certain extent. However, because the method is not aimed at enhancing the underwater image, there is a color deviation in the restoration result, and some details are not up to the standard. Nice enhancement. Figure 11 shows the comparison of layered and non-layered effects. It can be seen that the enhancement result loses some details, like the water bubbles in the first column, which further shows that the underwater image enhancement method based on structure-texture layering is more effective. it is good. In other words, the method mentioned in this paper achieves the best enhancement, not only effectively reducing noise, but also preserving good details and natural tones.

Claims (3)

1. An underwater image enhancement method based on structure-texture layering, characterized in that, first, color correction is performed through histogram equalization, and then the color-corrected image is decomposed into a low-frequency structure layer and a high-frequency texture layer , the noise is left in the texture layer, and then based on the proposed fog line model, the transmittance is accurately estimated from the structure layer without noise, and then enhanced, and then the texture layer is enhanced by the gradient residual minimization method. Appropriate scaling is used to reconstruct the enhanced structural layer, texture layer and refined edge mask into the final enhanced map, thereby solving the problems that cannot be handled by the existing technology; the specific steps are detailed as follows: 1) Perform color correction on the captured image: First, combined with the characteristics of the color distribution histogram of the underwater image, move the average color distribution of each color channel to the desired range, and then perform linear normalization, that is, use a simple The histogram equalization of the

Calculate the lower bound of the boundary as follows

and cap

where ^c∈ {R,G,B}, μc and ^σc are the mean and standard deviation of each color channel, respectively, λ is a hue parameter that effectively adjusts the distribution range, and then uses the truncation function chip() to convert The pixel value range of the image is truncated to [0, 255], resulting in a color-corrected image:

I ^c (x) is the pixel value of each channel of the color-corrected image; 2) Build a layered model: Define an underwater image with a dehazing model: I ^c (x)=J ^c (x)t(x)+A ^c (1-t(x))+E(x), (4) where x is the pixel value, J ^c (x) is the restored image, A ^c is the global background light, t(x) is the projection rate in water, define E(x) as the noise in the input image, the image enhancement The purpose is to recover J ^c (x) from the captured underwater image I ^c (x), which involves the estimation of transmittance and background light; 3) Structure-texture layering implementation: TV-L ¹ is used to separate the structure layer and the texture layer:

in,

is the gradient of the structural layer normalized by the 1-norm, ξ is the weight parameter controlling the smoothness, the structural layer I _s is solved by TV-L ¹ , and TV-L ¹ uses the total variation as a regular term to decompose the image Optimization method, the texture layer is calculated by It = II _{s ;} 4) Structural layer enhancement: Use the global background light scattering estimation method to estimate A:

in,

is the maximum position in the dark channel that reaches the structural layer, that is,

Ω(y) is a field of y; According to the dark channel prior theory, first assume that the transmittance in each window is constant, and define it as t ₀ . The value of A has been calculated by formula (9), so the minimum value is calculated twice on both sides of formula (6). , the estimated value of transmittance t(x) is obtained after shifting the items:

where ω is the correction factor; 5) Texture layer de-noising: Through the above TV-L ¹ decomposition, the noise in the captured image remains in the high-frequency texture layer, and the gradient residue minimization method is used to enhance the texture layer. The enhancement of this texture layer is achieved through the optimization formula (15) ) to get:

Among them, δ and η are the control weight parameters, which control the smoothness and fineness of the processing result. If η is too large or δ is too small, most of the texture of the enhanced image will be lost with the denoising process. Therefore, it is appropriate to select δ The value of and η has a great influence on the experimental results. Let Z=J-I, then the above optimization problem becomes two sub-problems, as follows:

This turns Equation (16) into a classic TV problem, and Equation (17) is approximated with a soft threshold; 6) Mask refinement: Use a binary mask M to separate the texture layer decomposed by TV-L ¹ into smooth areas and detail areas, and then delete the residual details in the smooth area of the enhanced low-frequency structure layer. Cosine transform coefficient to detect the similarity between blocks in the scene, to judge whether a region is smooth, define the _DCT coefficient of each 8 × 8 block in It as B, then, the similarity of each block in the scene details Sex is expressed by formula (18):

Among them, x, y are the coordinate positions in the DCT, calculate the square sum of all DCT coefficients except B _1,1 , B 1, ₂ and B ₂ , 1, and then use the threshold to judge the similarity of each block, define The initial edge mask of the texture layer is M, and the threshold κ is set to 0.1. When ρ is greater than κ, M of the block is 1, otherwise, M is 0, so that a rough binary edge information map is obtained. In order to obtain finer edge mask

The initial rough M is modified by the soft mapping method combined with the mapping _Labrador matrix generated by the structural layer Is:

where m and m' are the vector representations of M and M', L _s is the mapped _Labrador matrix generated from Is,

is the normalization parameter, which is set to 10 ⁵ in the text; 7) Reconstruction: Finally, we reconstruct the final enhancement processing result by formula (20):

where τ is the scaling factor, which introduces enhanced details, and is set as

and

They are the enhanced low-frequency structure layer and high-frequency texture layer, respectively. 2. The underwater image enhancement method based on structure-texture layering as claimed in claim 1, wherein the high frequency part of the image not only contains texture information, but also contains a large amount of noise in a low-light environment. The accuracy of transmittance estimation, firstly, the image is layered, and then the transmittance is estimated in the low-frequency structure layer, and the high-frequency texture layer is denoised. A picture is defined as J=J _s +J _t , J _s and J _t are the structure layer and texture layer of the image, respectively, then equation (4) is rewritten as: I ^c (x)=(J _s ^c (x)+J _t ^c (x))t(x)+A ^c (1-t(x))+E(x) =J _s ^c (x)t(x)+A ^c (1-t(x))+J _t ^c (x)t(x)+E(x).(5) The background light ^Ac is a relatively smooth signal, and the estimated transmittance t is also close to the ideal transmittance in J _s ^c . Therefore, the captured image can be written as: I=I _s +I _t , and we get: I _s ^c (x)=J _s ^c (x)t(x)+A ^c (1-t(x)) (6)

Among them, I _{sc is the low-frequency structure layer of the captured image, which contains most of the structure contours of the image, and I t c} ^is the high-frequency texture layer of the captured image, which contains most of the texture information and noise ^. Through this method, the transmittance _t (x) is estimated from Is (x) and is immune to noise. 3. the underwater image enhancement method based on structure-texture layering as claimed in claim 1 is characterized in that, the estimated value of transmittance t(x) is more refined processing step: formula (6) is converted into 3D In the coordinate system, a linear transmittance estimation value t'(x) in the three-dimensional coordinate system is obtained, and the background light A is used as the coordinate origin, and the formula (6) is changed into the following form: I _s0 (x)=t'(x) · J _s0 (x) (11) where I _s0 (x)=I _s (x)-A, J _s0 (x)=J _s (x)-A, therefore, the linear transmittance estimate for each pixel value is calculated according to the following formula:

Given a boundary value of transmittance

Next, by optimizing formula (13), a fine transmittance map is obtained:

where α and β are the weight parameters that control the data item and the smooth item, and σ(x) is

The standard deviation at each pixel value, N(x) represents the four-neighborhood pixel location around each pixel value in the structure layer, A and

After the value of is calculated, then, this enhanced structural layer can be obtained by Equation (14):

Images