Dual-scale nuclear power station underwater image color cast removing method based on nerve enhancement

Abstract

The invention discloses a dual-scale nuclear power station underwater image color cast removal method based on nerve enhancement, and belongs to the technical field of nuclear power station underwater image color cast removal. Based on the selective attenuation principle of light, a specific calculation mode of background light intensity A and improved transmissivity t is selected, the method is suitable for the underwater environment of the nuclear power station, the color cast problem of an underwater image can be well solved after a Koschmieder model is combined, the calculation proportion of the illumination intensity of a blue-green channel can be adjusted according to different water depths, and the applicability is high; noise in an underwater image of a nuclear power station and a morphological artifact problem caused by calculation are considered while color cast removal and refinement are performed on the image, and artifact removal, denoising and other processing are performed on the image by using a haze line averaging algorithm and a dual-scale GAN network.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear power plant underwater image decolorization, and in particular to a dual-scale nuclear power plant underwater image decolorization method based on neural enhancement. Background Art

[0002] The special optical properties of the underwater environment lead to the degradation of underwater image quality. Water molecules and various suspended matter and sediments in the water absorb and scatter light, thereby reducing the brightness and contrast of underwater images and causing blur and noise points in the image. In addition, artificial light sources added to meet specific shooting requirements will also affect image quality. Light of different wavelengths selectively attenuates when propagating in water, and red light with a longer wavelength has the weakest penetration ability. Therefore, underwater images are often dominated by blue-green tones, with obvious color cast problems, resulting in the visual quality of the original underwater images being difficult to meet the requirements of actual applications in underwater scenes.

[0003] These problems are more prominent in the underwater environment of nuclear power plants. The water used to cool and store nuclear fuel in the water pool of nuclear power plants has specific chemical properties such as high density and high purity, and is usually doped with boric acid to absorb neutrons. Although this high-purity water reduces some suspended particles, the high density of water and the added boric acid will further affect the propagation characteristics of light, exacerbating the color deviation problem of the image. In addition, nuclear fuel pools usually require strong light illumination to ensure the safety and accuracy of operation. The scattering and reflection of these strong light sources in water will increase the halo and noise points of the image. In addition, most nuclear power plants are semi-enclosed underwater, and image acquisition requires the assistance of artificial light sources. The above factors lead to the decline in the quality of the collected images, and the appearance of blur, distortion, and noise in the images, which further aggravates the color deviation problem of underwater images of nuclear power plants, making it difficult to accurately reflect the operating details of the equipment. Therefore, it is very important and necessary to develop feasible and effective underwater image decolorization and enhancement methods to improve the quality of images in the underwater environment of nuclear power plants, ensure operational safety, and monitor accuracy.

[0004] Existing methods for decolorizing underwater images mainly include the following three categories: methods based on image processing, methods based on physical models, and methods based on data-driven methods. (1) Methods based on image processing for decolorizing images mainly rely on traditional image processing techniques to remove color casts by adjusting the contrast, brightness, and color of the image. Common methods include histogram equalization, adaptive histogram equalization (CLAHE), color enhancement algorithms, etc. These methods do not consider the underwater optical transmission model, are relatively simple, and have high computational efficiency, but have some limitations: since they ignore the unique optical properties of the underwater environment, these methods may cause color distortion, and the image noise may be amplified during the enhancement process, affecting the image quality; unnatural colors may be introduced during the color adjustment process, affecting the visual effect. (2) Methods based on physical models use prior information and underwater imaging models to estimate the parameters required for image decolorization. They usually include steps such as atmospheric light estimation, transmittance calculation, and image restoration. Common methods include dark channel prior (DCP), etc. These methods obtain clear images by inverting the degradation process and can better restore the true color of the image. However, due to the reliance on underwater imaging parameter estimation, the applicability and stability are challenged in complex and changeable underwater environments. These methods are usually only applicable to specific underwater scenes and have high requirements for the accuracy of parameter estimation. (3) Data-driven methods use deep learning technology to train models with a large amount of underwater image data to achieve automatic decolorization. These methods mainly include generative adversarial networks (GANs) and convolutional neural networks (CNNs). These methods can generate high-quality decolorized images with good enhancement effects. However, this method has high data requirements and requires a large amount of underwater image data for training. In addition, the structure of the deep learning model is complex, the training time is long, and the demand for computing resources is high.

[0005] The above problems need to be solved urgently. Therefore, a dual-scale nuclear power plant underwater image decolorization method based on neural enhancement is proposed. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to solve the problem of severe color cast in underwater images of nuclear power plants as the depth increases, improve the quality of underwater images of nuclear power plants, meet the requirements of tasks such as monitoring and maintenance operation of underwater equipment of nuclear power plants, and provide a dual-scale nuclear power plant underwater image decolorization method based on neural enhancement.

[0007] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0008] S1: inputting an underwater degraded image of a nuclear power plant and preprocessing the image, wherein the underwater degraded image of the nuclear power plant is an underwater image of the nuclear power plant;

[0009] S2: Ignore the red channel and select the maximum value of the green and blue channels to estimate the initial value of the atmospheric light intensity;

[0010] S3: traverse each pixel of the image, calculate the ratio of the blue and green channel values ​​of the current pixel to the blue and green channel values ​​of the atmospheric light, and also ignore the red channel. Based on the exponential decay model, the propagation characteristics of light in the medium and the Beer-Lambert law, define the calculation formula of the transmittance t, and calculate the initial value of the transmittance according to the calculation formula of the transmittance t, and then obtain the initial transmission map t0 corresponding to the image;

[0011] S4: Reduce the morphological artifacts of the initial transmission image t0 through the haze line averaging algorithm to generate the transmission image t m ;

[0012] S5: The transmission map t is filtered using a guided filter based on the local weighted least squares error. m Perform smoothing and denoising to obtain a processed transmission map;

[0013] S6: Use generative adversarial networks to refine the atmospheric light and transmission maps and reduce or remove noise in the atmospheric light and transmission maps;

[0014] S7: constructing the Laplacian pyramid of the underwater image of the nuclear power plant, the Gaussian pyramid of the transmission map, and the Laplacian pyramid of the enhanced image, and performing decolorization, denoising, and refinement on each level of the Laplacian pyramid of the enhanced image to form the Laplacian pyramid of the enhanced image, and then reconstructing and enhancing the details of each level of the Laplacian pyramid of the enhanced image to obtain the final underwater enhanced image of the nuclear power plant;

[0015] S8: Perform global linear contrast stretching on the final nuclear power plant underwater enhanced image, convert the image from floating point format to unsigned 8-bit integer type, and realize decolorization processing of the nuclear power plant underwater image.

[0016] Furthermore, in step S1, the specific process of preprocessing is as follows:

[0017] S11: Change the input image format to floating point format;

[0018] S12: Read the input image size and channel number information, and assign the height, width and channel number information to parameters w, h and c;

[0019] S13: Set the pyramid layer number parameter l.

[0020] Furthermore, in step S2, the specific process is as follows:

[0021] S21: Obtain information such as the height, width, and number of color channels of the input image, and extract the maximum value of a specific channel by calculating the following formula:

[0022] max_blue=max(max(I(:,:,B)))

[0023] max_green=max(max(I(:,:,G)))

[0024] Among them, max_blue and max_green represent the maximum pixel values ​​of the blue channel and the green channel respectively, I(:,:,B) and I(:,:,G) are the pixel values ​​of the blue and green components of each pixel in the image respectively;

[0025] S22: Take the larger value as the initial value A of the global atmospheric light intensity, and ensure that its value is within the valid range. The formula is as follows:

[0026] A=max(0,max_blue,max_green);

[0027] A=min(max(A,0),255).

[0028] Furthermore, in step S3, the calculation formula of the transmittance t is defined as follows:

[0029]

[0030] Among them, t(i,j) is the transmittance, that is, the transmittance value at the pixel position (i,j); β is the attenuation coefficient of light in water; I(i,j,G) and I(i,j,B) are the pixel values ​​of the green and blue channels respectively, and A(G) and A(B) are the green and blue components of atmospheric light.

[0031] Furthermore, in step S4, the specific processing process is as follows:

[0032] S41: Perform gamma correction on the input image;

[0033] S42: converting the coordinate system into a coordinate system centered on the atmospheric light, and calculating the distance of each pixel point in the image relative to the atmospheric light;

[0034] S43: converting the coordinates of each pixel point in the image into a radius, elevation angle and azimuth angle in a spherical coordinate system;

[0035] S44: construct a histogram in spherical coordinates, and count the histogram bins and sub-bins to which each pixel belongs;

[0036] S45: Calculate the total number of sub-bins and record the first sub-bin index of each histogram bin;

[0037] S46: among all the pixels, according to the sub-boxes to which they belong, the transmittance and weight are cumulatively calculated; for each sub-box, the average transmittance is calculated using the accumulated transmittance and weight, and the average value is allocated to each pixel.

[0038] Furthermore, in step S5, the specific processing process of the guided filter based on the local weighted least squares error is as follows:

[0039] S51: Calculate the size, mean and variance of the local window, and then calculate the linear coefficient and offset;

[0040] S52: The image is guided by the linear coefficient and the offset to obtain a filtered output image, that is, a processed transmission map.

[0041] Furthermore, in step S6, the generator and the discriminator of the generative adversarial network are both constructed based on the Laplacian pyramid image model, and both are dual-scale structures.

[0042] Furthermore, in step S7, at the top layer of the Laplacian pyramid of the enhanced image, the enhanced image corresponding to the underwater image of the nuclear power plant is derived based on the Koschmieder model:

[0043] I(x)=(J(x)-A) / t(x)+A

[0044] Among them, J(x) is the underwater degraded image of the nuclear power plant, I(x) is the corresponding enhanced image, A is the initial value of the intensity of atmospheric light, and t(x) is the transmittance;

[0045] In the other levels of the Laplacian pyramid of the enhanced image except the top level, the enhanced image corresponding to the underwater image of the nuclear power plant is:

[0046]

[0047] Where l represents the level of the Gaussian pyramid of the underwater degraded image and transmission image of the nuclear power plant, J(l) represents the lth layer of the underwater degraded image of the nuclear power plant, and η is a constant.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. Based on the principle of selective attenuation of light, a specific background light intensity A and an improved calculation method of transmittance t are selected, which is suitable for the underwater environment of nuclear power plants. After combining with the Koschmieder model, the color cast problem of underwater images can be better solved, and the calculation ratio of the light intensity of the blue-green channel can be adjusted according to different water depths, which has strong applicability.

[0050] 2. While considering the decolorization and refinement of the image, the noise in the underwater image of the nuclear power plant and the morphological artifacts caused by calculation are also considered. The haze line averaging algorithm and the dual-scale GAN network are used to remove artifacts and denoise the image.

[0051] 3. The model-based method and the data-driven method are combined to well preserve the high-quality visual characteristics of the model-based method, and the model-based estimation and data-driven refinement form neural enhancement. Theoretical analysis shows that the proposed neural enhancement framework converges faster than the corresponding data-driven method. And the results show that the algorithm proposed in this invention is superior to the existing data-driven underwater image decolorization algorithm in terms of evaluation indicators such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of a method for decolorizing a dual-scale nuclear power plant underwater image based on neural enhancement in Embodiment 2 of the present invention;

[0053] Figure 2 1 is a schematic diagram of the structure of the generative adversarial network and related modules in the second embodiment of the present invention, wherein (a) is the generative adversarial network, (b) is the dual attention module, (c) is the recursive residual group module, and (d) is the specific network structure of the G1, G2, and G3 modules in the generative adversarial network;

[0054] Figure 3 is the underwater degraded image of the nuclear power plant input in the second embodiment of the present invention;

[0055] Figure 4 This is the final underwater enhanced image of the nuclear power plant in the second embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0057] Embodiment 1

[0058] This embodiment provides a technical solution: a method for removing color deviation of a dual-scale underwater image of a nuclear power plant based on neural enhancement, comprising the following steps:

[0059] S1: Estimate the initial value A of the global atmospheric light intensity from the underwater degraded image of the nuclear power plant, that is, the intensity of the underwater background light.

[0060] Based on the law of selective attenuation of light and the definition and estimation method of atmospheric light, first of all, the propagation of light in water is affected by absorption and scattering, and different wavelengths of light attenuate to different degrees. Red light (long wavelength) attenuates most significantly in water and has the weakest penetration ability. Green and blue light (short wavelength) have strong penetration ability. Therefore, in underwater environments, these two colors of light dominate. The initial value of the intensity of the atmospheric light of the underwater image is estimated by selecting the maximum value of the green and blue channels. Since red light attenuates faster in underwater environments, the influence of the red channel is ignored, and the light intensity of the blue and green channels is focused on.

[0061] The estimation process of the initial value of strength A is as follows:

[0062] Get the height, width, and number of color channels of the input image, and extract the maximum value of a specific channel by the following calculation:

[0063] max_blue=max(max(I(:,:,B)))

[0064] max_green=max(max(I(:,:,G)))

[0065] Among them, max_blue and max_green represent the maximum pixel values ​​of the blue channel and the green channel respectively, and I(:,:,B) and I(:,:,G) are the pixel values ​​of the blue and green components of each pixel in the image respectively;

[0066] Then take the larger value as the initial value A of the global atmospheric light intensity, and ensure that its value is within the valid range:

[0067] A=max(0,max_blue,max_green);

[0068] A=min(max(A,0),255).

[0069] S2: Calculate the initial transmittance.

[0070] Traverse each pixel of the input image and calculate the ratio of the blue and green channel values ​​of the current pixel to the blue and green channel values ​​of the atmospheric light. Based on the exponential decay model, the propagation characteristics of light in the medium and the Beer-Lambert law, the calculation formula of the transmittance t is defined:

[0071]

[0072] Among them, t(i,j) is the transmittance, that is, the transmittance value at the pixel position (i,j); β is the attenuation coefficient of light in water; I(i,j,G) and I(i,j,B) are the pixel values ​​of the green and blue channels respectively, and A(G) and A(B) are the green and blue components of atmospheric light. According to this formula, the initial transmission map t0 corresponding to the entire image is calculated;

[0073] S3: Reduce the morphological artifacts of the initial transmission image t0 through the haze line averaging algorithm to generate the transmission image t m .

[0074] The specific steps include: (1) image preprocessing: gamma correction is performed on the input underwater image of the nuclear power plant to correct uneven illumination and enhance image contrast; (2) distance calculation relative to atmospheric light: the coordinate system is converted into a coordinate system centered on atmospheric light, and the distance of each pixel in the image relative to atmospheric light is calculated; (3) conversion to a spherical coordinate system: the coordinates of each pixel in the image are converted into the radius, elevation angle and azimuth angle in the spherical coordinate system; (4) histogram statistics: a histogram in spherical coordinates is constructed, and the histogram bins and sub-bins to which each pixel belongs are counted; (5) total number of sub-bins is calculated: the total number of sub-bins is calculated, and the first sub-bin index of each histogram bin is recorded; (6) transmittance averaging: the transmittance is averaged according to the selected method to obtain a smoother and more reasonable transmittance estimate.

[0075] S4: Data-driven refinement structure, using generative adversarial networks (GAN) to refine atmospheric light and transmission maps. The generator and discriminator are both dual-scale, based on the Laplacian pyramid image model. Because the atmospheric light and transmission maps estimated by model-based methods often rely on specific assumptions, and due to the interference of complex scenes and the input image itself may contain noise, the estimated atmospheric light and transmission maps often contain noise, so a dual-scale refinement method is proposed to reduce / remove noise.

[0076] S5: Construct the Laplacian pyramid of the underwater image, the Gaussian pyramid of the transmission image, and the Laplacian pyramid of the enhanced image, and perform calculations such as decolorization, denoising, and refinement on each level of the Laplacian pyramid of the enhanced image. Based on the Koschmieder model:

[0077] J(x)=I(x)t(x)+A(1-t(x))

[0078] Among them, J(x) is the underwater degraded image of the nuclear power plant, I(x) is the corresponding enhanced image, A is the intensity of atmospheric light (here it is the intensity of underwater background light), and t(x) is the transmittance, which reflects the attenuation of light when it propagates in water. It indicates the degree of absorption and scattering of light by water, which usually varies with the depth and color of the water. According to the inversion degradation process, we can get:

[0079] I(x)=(J(x)-A) / t(x)+A

[0080] Thus, the corresponding enhanced image calculation method can be obtained. In the top layer of the Laplace pyramid, the corresponding enhanced image can be obtained according to the above formula, which is used as the top layer of the enhanced image Laplace pyramid. In other layers except the top layer, the calculation formula of the enhanced image is:

[0081]

[0082] Where l represents the level of the Gaussian pyramid of the underwater degraded image and transmission map of the nuclear power plant, J(l) represents the lth layer of the underwater degraded image of the nuclear power plant, and η is a constant used to adapt to different image degradation levels and will decrease as the degradation level increases. Here it is 1 / 4, which is suitable for the case of low degradation level, and is often set to 1 / 8 when the degradation level is high. The top layer of the Laplacian pyramid of the enhanced image is combined with other levels to form a complete Laplacian pyramid of the enhanced image, and then the Laplacian pyramid is used to reconstruct and enhance the details of the image to obtain the final enhanced underwater image of the nuclear power plant.

[0083] Embodiment 2

[0084] like Figure 1 As shown, the present invention provides a dual-scale nuclear power plant underwater image decolorization method based on neural enhancement, and the specific steps of the method are as follows:

[0085] S1: Input underwater image I of the nuclear power plant, i.e., underwater degraded image of the nuclear power plant, see Figure 3 ;

[0086] S2: preprocess the underwater image I and convert it into double-precision floating point format, obtain the height, width and number of channels of the preprocessed underwater image I and assign them to parameters w, h and c, and set the pyramid layer number parameter l;

[0087] S3: Select the maximum value of the green and blue channels to estimate the initial value A of the global atmospheric light of the underwater image, and ignore the influence of the red channel to adapt to the special underwater environment of the nuclear power plant;

[0088] S4: Traverse each pixel of the input image, calculate the ratio of the blue and green channel values ​​of the current pixel to the blue and green channel values ​​of the atmospheric light, and also choose to ignore the red channel; based on the exponential decay model, the propagation characteristics of light in the medium and the Beer-Lambert law, define the calculation formula of the transmittance t applicable to the underwater environment:

[0089]

[0090] Among them, t(i,j) is the transmittance, that is, the transmittance value at the pixel position (i,j); β is the attenuation coefficient of light in water; I(i,j,G) and I(i,j,B) are the pixel values ​​of the green and blue channels respectively, and A(G) and A(B) are the green and blue components of atmospheric light; according to this formula, the initial transmission map t0 corresponding to the entire image is obtained.

[0091] S5, reduce the morphological artifacts of the initial transmission image t0 through the haze line average algorithm to generate a transmission image t m .

[0092] The specific steps include: (1) Image preprocessing: gamma correction is performed on the input underwater image of the nuclear power plant to correct uneven illumination and enhance image contrast; (2) Calculation of distance relative to atmospheric light: The coordinate system is converted into a coordinate system centered on atmospheric light, and the distance of each pixel in the image relative to atmospheric light is calculated; (3) Conversion to spherical coordinate system: The coordinates of each pixel in the image are converted into the radius, elevation and azimuth in the spherical coordinate system; (4) Histogram statistics: Construct a histogram in spherical coordinates and count the histogram bins and sub-bins to which each pixel belongs; (5) Calculation of the total number of sub-bins: Calculate the total number of sub-bins and record the first sub-bin index of each histogram bin; (6) Transmittance average: Among all pixels, the transmittance and weight are accumulated according to the sub-bin to which they belong; For each sub-bin, the accumulated transmittance and weight are used to calculate its average transmittance, and the average value is assigned to each pixel. Through the accumulation and weighted average method, the transmittance estimation is effectively smoothed, local noise and artifacts are reduced, and the transmittance map is smoother and more reasonable.

[0093] S6: The image (i.e., the transmission map t m ) for smoothing and denoising. The basic steps are: calculate the size, mean and variance of the local window, and then calculate the linear coefficient and offset, and use the linear coefficient and offset to guide the image to obtain the filtered output image. While retaining the edges and details of the image, it can effectively remove noise.

[0094] S7: Data-driven refinement structure, using generative adversarial networks (GAN) to refine atmospheric light and transmission maps. Figure 2 As shown in (a), GAN is used to refine the atmospheric light and transmission maps. Since it is difficult to refine the transmission map, the generator of atmospheric light is simpler than that of the transmission map. In addition, a dual attention module (DAB) and a recursive residual group module (RRG) are added to the generative adversarial network (GAN) to further refine the underwater image of the nuclear power plant after decolorization.

[0095] The dual attention module (DAB) helps the network focus on important image areas and features more effectively by introducing spatial attention mechanism and channel attention mechanism, thereby improving the effect of image restoration. Figure 2 As shown in (b), the spatial attention mechanism: perform global average pooling and global maximum pooling on the feature map to obtain two single-channel feature maps, then splice the two feature maps, and generate a weight map through the convolution layer; finally, multiply the input feature map and the weight map element by element to highlight the key spatial area. Channel attention mechanism: perform global maximum pooling on the feature map in the spatial dimension, adjust the number of channels through 1×1 convolution, PRelu activation function and the second 1×1 convolution, generate a channel attention weight vector, and fix the weight value between [0,1] through the Sigmoid function; finally, multiply the calculated channel weight vector and the input feature map element by element to enhance the feature response of important channels and suppress unimportant channels. Finally, the processed features of the two are spliced, and the resulting feature map not only focuses on important areas in space, but also highlights key features in channels, improving the network's image refinement ability.

[0096] The recursive residual group module (RRG) consists of multiple dual attention modules (DAB), such as Figure 2 As shown in (c), skip connections are introduced to achieve more efficient gradient propagation. Through feature extraction operations at different scales, information at all levels in the image can be captured. Through multi-level residual connections, global and local feature information can be effectively combined. This is conducive to enhancing image contrast and restoring image details.

[0097] In this embodiment, the generator and discriminator of the generative adversarial network are both dual-scale and based on the Laplacian pyramid image model. The main function is to reduce or remove the noise in the atmospheric light and transmission map estimated by the model-based method. In addition, during the network training phase, only 1500 sets of 256×256 decolorized image pairs were used, including underwater degraded images of nuclear power plants and corresponding decolorized enhanced images. The training data is significantly reduced compared to the existing data-driven decolorization algorithms.

[0098] S8: Construct the Laplacian pyramid of the underwater image of the nuclear power plant, the Gaussian pyramid of the transmission image, and the Laplacian pyramid of the enhanced image, and perform calculations such as decolorization, denoising, and refinement on each level of the Laplacian pyramid of the enhanced image. At the coarsest level (i.e., the top level of the Laplacian pyramid), based on the Koschmieder model:

[0099] J(x)=I(x)t(x)+A(1-t(x))

[0100] By deducing the model, the enhanced image corresponding to the underwater degraded image is obtained as follows:

[0101] I(x)=(J(x)-A) / t(x)+A

[0102] Among them, J(x) is the underwater degraded image of the nuclear power plant (i.e., the underwater image I of the nuclear power plant in step S1 above), I(x) is the corresponding enhanced image, A is the initial value of the atmospheric light intensity (here, the underwater background light intensity), and t(x) is the transmittance that describes how the atmospheric light reaches the camera, reflecting the attenuation of light propagating in water. It represents the degree of absorption and scattering of light by water, which usually varies with the depth and color of the water. In each pyramid level except the top level, the calculation method of the underwater enhanced image of the nuclear power plant is defined as:

[0103]

[0104] Where l represents the lth layer of the underwater degraded image or the transmission image Gaussian pyramid, J(l) represents the lth layer of the underwater degraded image of the nuclear power plant, and η is a constant used to adapt to different image degradation levels and will decrease as the degradation level increases. Here it is set to 1 / 4. And there are two important cases in this equation, as follows:

[0105] (1): When the transmittance t(l) is much smaller than the constant η, then is approximately equal to 1, then the above formula can be derived as:

[0106]

[0107] When l = 1:

[0108] I(l)=J(l)

[0109] When l = 2:

[0110]

[0111] As the intensity of detail enhancement and noise processing gradually weakens from the bottom to the top of the pyramid, and gradually decreases from the top to the bottom Therefore, it can ensure the detailed processing of high-resolution layer noise and the smooth transition of low-resolution layer images to prevent noise amplification.

[0112] (2): When the transmittance t(l) is much larger than the constant η, then is approximately equal to 0, the above formula can be derived as:

[0113]

[0114] In this case, the underwater enhanced image of the nuclear power plant is close to Objects closer to the camera are better recovered and have enhanced details due to their higher transmittance.

[0115] Finally, the information of each layer of the Laplacian pyramid of the enhanced image is merged layer by layer to generate the final underwater enhanced image of the nuclear power plant. This adaptive processing method of the algorithm combines the spatial characteristics of the image to ensure that the details of each area can be properly enhanced or protected, and based on the calculation method of background light and transmittance, it effectively completes the task of decoloring the underwater image of the nuclear power plant.

[0116] S9: Perform global linear contrast stretching on the enhanced underwater image of the nuclear power plant, and adjust the image by combining the set cropping ratio with the adjust function to improve the overall contrast of the image. Convert the image from floating point format to an unsigned eight-bit integer type (0-255 range), and compare it with the intensity value A of the background light, the transmission map t (i.e., the transmission map after the generative adversarial network refinement), and the enhanced image (see Figure 4 ) are saved together into the structure ouput_Mat and output.

[0117] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A dual-scale nuclear power plant underwater image decolorization method based on neural enhancement, characterized in that: The following steps are involved: S1: inputting an underwater degraded image of a nuclear power plant and preprocessing the image, wherein the underwater degraded image of the nuclear power plant is an underwater image of the nuclear power plant; S2: Ignore the red channel and select the maximum value of the green and blue channels to estimate the initial value of the atmospheric light intensity; S3: traverse each pixel of the image, calculate the ratio of the blue and green channel values ​​of the current pixel to the blue and green channel values ​​of the atmospheric light, and also ignore the red channel. Based on the exponential decay model, the propagation characteristics of light in the medium and the Beer-Lambert law, define the calculation formula of the transmittance t, and calculate the initial value of the transmittance according to the calculation formula of the transmittance t, and then obtain the initial transmission map t0 corresponding to the image; S4: Reduce the morphological artifacts of the initial transmission image t0 through the haze line averaging algorithm to generate the transmission image t m ; S5: The transmission map t is filtered using a guided filter based on the local weighted least squares error. m Perform smoothing and denoising to obtain a processed transmission map; S6: Use generative adversarial networks to refine the atmospheric light and transmission maps and reduce or remove noise in the atmospheric light and transmission maps; S7: constructing the Laplacian pyramid of the underwater image of the nuclear power plant, the Gaussian pyramid of the transmission map, and the Laplacian pyramid of the enhanced image, and performing decolorization, denoising, and refinement on each level of the Laplacian pyramid of the enhanced image to form the Laplacian pyramid of the enhanced image, and then reconstructing and enhancing the details of each level of the Laplacian pyramid of the enhanced image to obtain the final underwater enhanced image of the nuclear power plant; S8: Perform global linear contrast stretching on the final nuclear power plant underwater enhanced image, convert the image from floating point format to unsigned 8-bit integer type, and realize decolorization processing of the nuclear power plant underwater image.

2. According to the neural enhancement-based dual-scale nuclear power plant underwater image decolorization method of claim 1, it is characterized in that: In step S1, the specific process of preprocessing is as follows: S11: Change the input image format to floating point format; S12: Read the input image size and channel number information, and assign the height, width and channel number information to parameters w, h and c; S13: Set the pyramid layer number parameter l.

3. According to the neural enhancement-based dual-scale nuclear power plant underwater image decolorization method of claim 1, it is characterized in that: In step S2, the specific process is as follows: S21: Obtain the height, width and number of color channels of the input image, and extract the maximum value of a specific channel by the following calculation: max_blue=max(max(I(:,:,B))) max_green=max(max(I(:,:,G))) Among them, max_blue and max_green represent the maximum pixel values ​​of the blue channel and the green channel respectively, I(:,:,B) and I(:,:,G) are the pixel values ​​of the blue component and the green component of each pixel in the image respectively; S22: Take the larger value as the initial value A of the global atmospheric light intensity, and ensure that its value is within the valid range. The formula is as follows: A=max(0,max_blue,max_green); A=min(max(A,0),255).

4. According to the neural enhancement-based dual-scale nuclear power plant underwater image decolorization method of claim 1, it is characterized in that: In step S3, the calculation formula of the transmittance t is defined as follows: Among them, t(i,j) is the transmittance, that is, the transmittance value at the pixel position (i,j); β is the attenuation coefficient of light in water; I(i,j,G) and I(i,j,B) are the pixel values ​​of the green and blue channels respectively, and A(G) and A(B) are the green and blue components of atmospheric light.

5. The method for decolorizing underwater images of a nuclear power plant based on neural enhancement in dual scales according to claim 1, characterized in that: In step S4, the specific processing process is as follows: S41: Perform gamma correction on the input image; S42: converting the coordinate system into a coordinate system centered on the atmospheric light, and calculating the distance of each pixel point in the image relative to the atmospheric light; S43: converting the coordinates of each pixel point in the image into a radius, elevation angle and azimuth angle in a spherical coordinate system; S44: construct a histogram in spherical coordinates, and count the histogram bins and sub-bins to which each pixel belongs; S45: Calculate the total number of sub-bins and record the first sub-bin index of each histogram bin; S46: among all the pixels, the transmittance and the weight are cumulatively calculated according to the subbox to which they belong; for each subbox, the average transmittance is calculated using the accumulated transmittance and weight, and the average is allocated to each pixel.

6. The method for decolorization of dual-scale underwater images of nuclear power plants based on neural enhancement according to claim 1, characterized in that: In step S5, the specific processing process of the guided filter based on the local weighted least squares error is as follows: S51: Calculate the size, mean and variance of the local window, and then calculate the linear coefficient and offset; S52: The image is guided by the linear coefficient and the offset to obtain a filtered output image, that is, a processed transmission map.

7. The method for decolorization of underwater images of a nuclear power plant based on neural enhancement in dual scales according to claim 1, characterized in that: In step S6, the generator and the discriminator of the generative adversarial network are both constructed based on the Laplacian pyramid image model, and both are dual-scale structures.

8. The method for decolorizing underwater images of a nuclear power plant based on neural enhancement in dual scales according to claim 7, characterized in that: In step S7, at the top layer of the Laplacian pyramid of the enhanced image, the enhanced image corresponding to the underwater image of the nuclear power plant is derived based on the Koschmieder model: I(x)=(J(x)-A) / t(x)+A Among them, J(x) is the underwater degraded image of the nuclear power plant, I(x) is the corresponding enhanced image, A is the initial value of the intensity of atmospheric light, and t(x) is the transmittance; In the other levels of the Laplacian pyramid of the enhanced image except the top level, the enhanced image corresponding to the underwater image of the nuclear power plant is: Where l represents the level of the Gaussian pyramid of the underwater degraded image and transmission image of the nuclear power plant, J(l) represents the lth layer of the underwater degraded image of the nuclear power plant, and η is a constant.

Images