CN114117908A - High-precision ASI sea ice density inversion algorithm for data correction based on CGAN - Google Patents

Abstract

The invention provides a high-precision ASI sea ice density inversion algorithm for data correction based on CGAN, belonging to the technical field of satellite remote sensing and comprising the following steps: s1, data screening: finding a relatively stable relation between 89GHz data and 36GHz data which are not interfered by external environments such as cloud, water vapor and the like or are slightly interfered, and screening out 89GHz data which are greatly interfered; s2, data correction: the generation network and the judgment network of the CGAN correct 89GHz data with large interference by using a stable relation between 36GHz data with high reliability and 89GHz data which is not interfered by external environments such as cloud, water vapor and the like or has small interference through countermeasure training; and S3, inverting the sea ice density by adopting an ASI sea ice density algorithm based on the corrected 89GHz data. The invention can greatly reduce the error caused by atmosphere by effectively correcting the sea ice density of the mixed pixels.

Description

High-precision ASI sea ice density inversion algorithm for data correction based on CGAN

Technical Field

The invention relates to the technical field of satellite remote sensing, in particular to a high-precision ASI sea ice density inversion algorithm for data correction based on CGAN.

Background

The change of the sea ice in the polar region is closely related to global atmospheric change, ocean circulation change and the like, and plays an important role in researching global climate change. Sea Ice Concentration (SIC) is one of the important parameters for studying the characteristics of sea ice space-time variation.

Many inversion algorithms have been proposed over the last several decades for estimating more accurate SIC values. The comma (1986,1995) proposes bootstrap algorithm, which mainly uses the change of sea ice emissivity with frequency band and different physical properties of sea water to invert the sea ice density. Cavelieri et al (1984, 1991) propose an NT algorithm that establishes an inversion equation for sea ice density based on differences in radiant brightness temperature in combination with polarization gradient rate and spectral gradient rate. In 2015, Liu et al (2015) propose an algorithm for calculating the south Pole sea ice intensity by using a fully constrained least squares algorithm on the basis of an NT algorithm. Cavalieri et al (1991) developed the NASA Team algorithm using SSM/I light temperature data, which was used for inversion of annual and perennial ice intensity. Markus et al (2000) added 89GHz vertical and horizontal polarization brightness temperature data to the NASA Team, suggesting the NASA Team 2 algorithm. Lomax et al (1995) developed the Lomax algorithm to calculate sea ice intensity using SSM/I85.5 GHz data. Hao (2015) improves the existing NASA Team algorithm by introducing AMSR-E6.9 GHz data, and improves the calculation accuracy of ice results for years. Kern and Heygster (2001) proposed the SEA LION algorithm that inverts for SEA ice density using polarization values in the 85.5GHz band and corrects for polarization differences using a radiative transfer model and atmospheric data obtained from numerical weather prediction models and/or SSM/I measurements. The ASI algorithm (ARTIST sea ice algorithm) was generated in the 1998 "study of Arctic radiation and turbulence exchange" project, Svendsen et al, based on the concept of "polarization corrected temperature" and uses data near the 90GHz band to invert the sea ice concentration (Spencer et al, 1989; Svendsen et al, 1987). The algorithm proposed by Svendsen et al was modified by Kaleschke et al to use higher resolution SSM/I85 GHz data to perform a mesoscale numerical simulation of the atmospheric boundary layer at the arctic sea ice edge (Kaleschke et al, 2001). One advantage of this algorithm is that it does not require additional data input, can perform result inversion directly from measured bright temperature data, and has similar results to sea ice density algorithms using other channels (Kern,2004), Spreen et al (2008) apply the ASI algorithm to AMSR-E data and obtain the corresponding sea ice density calculation formula, compared to other algorithms using 85GHz band data. Wang (2009) proposes a method for calculating the perennial ice density based on different characteristics between first year ice, perennial ice and seawater in the 89GHz band. Su et al (2013) performed a series of experiments, including interpolation algorithm "tie points" based on the ASI algorithm and weather filters. Zhang et al (2012) proposes a method for calculating sea ice density using multiband and dual polarization based on sea ice and sea water radiation characteristics.

The resolution and inversion algorithm of the satellite data is critical to accurately provide sea ice density. The Microwave data product with the highest resolution is currently a 6.25km resolution sea ice density grid product inverted by ASI algorithm based on 89GHz high-frequency data of a Microwave Scanning Radiometer AMSR-E (Advanced Microwave Scanning Radiometer-Earth Observing System) (Spreen et al, 2008). Although the ASI algorithm has advantages, compared with low-frequency data, 89GHz band data is more influenced by the density of atmospheric cloud liquid water, water drops, water vapor and ice surface snow particles, particularly in a thin ice region, when the liquid water content in the cloud is high or the cloud passes by a cyclone, the inversion result has larger error; meanwhile, the density of the accumulated snow particles on the surface of the sea ice is very sensitive; therefore, the ASI algorithm requires weather filter processing (Spreen et al, 2008). Although some errors can be eliminated using a weather filter, the intensity of the pixels cannot be modified when blending. In order to obtain a more accurate sea ice density result, 89GHz data which is greatly interfered by external environments such as cloud, water vapor and the like needs to be screened and corrected.

Disclosure of Invention

In view of this, the present invention provides a high-precision ASI sea ice density inversion algorithm for data correction based on CGAN, which can greatly reduce the error caused by the atmosphere by effectively correcting the sea ice density of the mixed pixels.

In order to solve the above technical problems, the present invention provides a high-precision ASI sea ice density inversion algorithm for data correction based on CGAN, comprising the steps of:

s1, data screening: finding a relatively stable relation between 89GHz data and 36GHz data which are not interfered by external environments such as cloud, water vapor and the like or are slightly interfered, and screening out 89GHz data which are greatly interfered;

s2, data correction: the generation network and the judgment network of the CGAN correct 89GHz data with large interference by using a stable relation between 36GHz data with high reliability and 89GHz data which is not interfered by external environments such as cloud, water vapor and the like or has small interference through countermeasure training;

the model function of CGAN is shown in (3):

wherein, x is data influenced by external environment, y is additional information, z is input random noise, G (z | y) is data which is not influenced by external environment and is input to the generation network of the CGAN under the influence of external environment, and D (G (z | y)) is the probability of judging the input data to be false by the judgment network; since the goal of the CGAN's generation network is to make the generated data as close as possible to data that is not affected by the external environment, the penalty function is set to 1-D (G (z | y)) to ensure that the discrimination network outputs false image probabilities as small as possible, and the goal of the discrimination network is to improve the ability to discriminate differences in the input data, so the larger D (x | y) the better, while the smaller the noise impact is desired to be the better, the penalty function is set to D (x | y) +1-D (G (z | y)), and so on

Indicating the progress of the game;

the CGAN model is trained as follows:

(1) before training, adjusting the data sets, such as rotating and translating, increasing the number of the data sets, and then carrying out normalization processing on the training set and the test set;

(2) inputting the training set into a generating network, and then carrying out continuous batch normalization + convolution + ReLU + pooling operation to complete downsampling operation;

(3) performing continuous operation of deconvolution, BN and an activation function on the feature map obtained by the down-sampling to finish the up-sampling;

(4) in the same network layer, the output characteristic mapping of the down sampling and the up sampling are connected; fusing the output characteristic diagrams of down-sampling and up-sampling of the next neural network layer from the top layer neural network layer to the bottom layer neural network layer in different network layers, continuously connecting the fused output characteristic diagrams by the up-sampling characteristic diagram of the next neural network layer, and continuously iterating until the next layer has no corresponding up-sampling, and then obtaining the mapping relation between 89GHz data and 36GHz data with high reliability,

(5) inputting the test set data, namely the relationship between the undisturbed 89GHz data and the 36GHz data, and the mapping relationship between the 89GHz data and the highly reliable 36GHz data obtained in the step (4) into a discrimination network;

(6) then, continuous batch normalization + convolution + ReLU + pooling operation is carried out to complete down-sampling operation;

(7) finally, judging the result obtained by the judgment network by using the cross entropy, outputting the corrected 89GHz data if the loss function reaches the minimum value, otherwise, returning to the step (2), and repeating the steps until the loss function reaches the minimum value;

and S3, inverting the sea ice density by adopting an ASI sea ice density algorithm based on the corrected 89GHz data.

Furthermore, the method for screening 89GHz data with large interference is concretely,

(1) under the condition of clear weather, a polarization ratio scatter diagram is drawn by taking a 36GHz data polarization ratio as an abscissa and an 89GHz data polarization ratio as an ordinate;

wherein the polarization ratio formula is shown as (1),

wherein TBV and TBH are respectively vertical polarization bright temperature and horizontal polarization bright temperature;

(2) in the polarization ratio scattergram, the abscissa is equally divided into several intervals, and the mean value and standard deviation of the ordinate in each interval are calculated, and then, a least squares best fit curve is drawn using the mean value minus two times the standard deviation in each interval, the curve being approximated by a quadratic equation, as shown in (2),

PR₈₉＝a(PR₃₆)²-bPR₃₆+c (2)

wherein PR89 is the polarization ratio of 89GHz data, PR36 is the polarization ratio of 36GHz data, and a, b and c are constants.

Furthermore, because the characteristics of different network layers can be integrated by the Unet network through a jump connection mode, the denoising performance is improved, and the Unet network has stronger self-adaptability and can effectively retain structural information of images, the Unet network is adopted as a generation network G;

wherein, Unet includes input layer, convolution layer, pooling layer, activation layer and output layer;

specifically setting the kernel of the pooling layer to be 2 × 2, setting the size of the convolution filter to be 3 × 3, and selecting the ReLU as an activation function;

the discrimination network D has the function of discriminating two groups of mapping relations, the discrimination network D adopts a CNN network, the image is subjected to feature extraction through continuous down-sampling layers after being input into the discrimination network, the sizes of convolution kernels in the down-sampling layers are all 4 multiplied by 4, the steps are all 2, in order to prevent the situation that the gradient disappears when the parameters are updated by using a gradient descent method, the convolution operation of the down-sampling layers is subjected to normalization processing, nonlinear mapping is carried out by using a ReLU activation function, finally, the value of each pixel is output by using a Sigmoid function in the discrimination layer, and the final loss is obtained by using cross entropy calculation.

The core idea of the CGAN model is as follows: nash balance is achieved through games of the generating network and the judging network, namely the judging network and the generating network achieve better results. The purpose of generating a network is to generate data that is nearly unaffected by the external environment, to improve the generation capability and to reduce the discrimination capability of the discrimination network.

The technical scheme of the invention has the following beneficial effects:

the invention provides a remote sensing image data correction method for generating a countermeasure network (CGAN) based on improved conditions. Through countermeasure training of a generation network and a judgment network of the CGAN, the 89GHz data with large interference is corrected by utilizing the stable relation between the 36GHz data with high reliability and the 89GHz data which is not interfered by external environments such as cloud, water vapor and the like or is small in interference.

On the basis, the ASI sea ice density algorithm is adopted to invert the sea ice density. The method makes full use of the 36GHz brightness temperature information with high reliability, can make up for the confusion of sea ice and sea water brightness temperature information in the observation process, improves the inversion accuracy of SIC, and can realize high spatial resolution at the same time.

The result shows that compared with the result of the ASI sea ice density inversion algorithm, the CGAN-based sea ice density inversion algorithm is feasible, and compared with the sea ice density obtained by Landsat data, the CGAN-based sea ice density inversion algorithm obviously improves the inversion accuracy of the sea ice density.

Drawings

FIG. 1 is a flowchart of the 89GHz affected data correction based on the improved CGAN model of the present invention;

FIG. 2 is a plot of polarization ratio scatter plots in an embodiment of the present invention;

FIG. 3 is a chart of sea ice concentration results obtained in an embodiment of the present invention;

FIG. 4 is a graph of sea ice concentration results obtained in comparative example 1 of the present invention;

FIG. 5 is a result of sea ice density for the selected area of FIG. 3 in an embodiment of the present invention;

FIG. 6 is a plot of the sea ice concentration results for the selected regions of FIG. 4 in comparative example 1 of the present invention;

fig. 7 is a result of sea ice distribution obtained from Landsat8 OLI data using reflectance thresholding in comparative example 3 of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the drawings of the embodiments of the present invention. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the described embodiments of the invention, are within the scope of the invention.

As shown in fig. 1-3, 5: the embodiment provides a high-precision ASI sea ice density inversion algorithm for data correction based on CGAN, which comprises the following steps:

s1, data screening: finding a relatively stable relation between 89GHz data and 36GHz data which are not interfered by external environments such as cloud, water vapor and the like or are slightly interfered, and screening out 89GHz data which are greatly interfered;

wherein, the method for screening 89GHz data with large interference is concretely,

(1) under the condition of clear weather, a polarization ratio scatter diagram is drawn by taking a 36GHz data polarization ratio as an abscissa and an 89GHz data polarization ratio as an ordinate;

wherein the polarization ratio formula is shown as (1),

wherein TBV and TBH are respectively vertical polarization bright temperature and horizontal polarization bright temperature;

(2) in the polarization ratio scattergram, the abscissa is equally divided into several intervals, and the mean value and standard deviation of the ordinate in each interval are calculated, and then, a least squares best fit curve is drawn using the mean value minus two times the standard deviation in each interval, the curve being approximated by a quadratic equation, as shown in (2),

PR₈₉＝a(PR₃₆)²-bPR₃₆+c (2)

wherein PR89 is the polarization ratio of 89GHz data, PR36 is the polarization ratio of 36GHz data, and a, b and c are constants;

the plot of the polarization ratio is shown in fig. 2, and is specifically obtained, where a is 3.5504, b is 0.1876, and c is 0.0061.

S2, data correction: the generation network and the judgment network of the CGAN correct 89GHz data with large interference by using a stable relation between 36GHz data with high reliability and 89GHz data which is not interfered by external environments such as cloud, water vapor and the like or has small interference through countermeasure training;

the model function of CGAN is shown in (3):

wherein, x is data influenced by external environment, y is additional information, z is input random noise, G (z | y) is data which is not influenced by external environment and is input to the generation network of the CGAN under the influence of external environment, and D (G (z | y)) is the probability of judging the input data to be false by the judgment network; since the goal of the CGAN's generation network is to make the generated data as close as possible to data that is not affected by the external environment, the penalty function is set to 1-D (G (z | y)) to ensure that the discrimination network outputs false image probabilities as small as possible, and the goal of the discrimination network is to improve the ability to discriminate differences in the input data, so the larger D (x | y) the better, while the smaller the noise impact is desired to be the better, the penalty function is set to D (x | y) +1-D (G (z | y)), and so on

Indicating the progress of the game;

the CGAN model is trained as follows:

(1) before training, adjusting the data sets, such as rotating and translating, increasing the number of the data sets, and then carrying out normalization processing on the training set and the test set;

(2) inputting the training set into a generating network, and then carrying out continuous batch normalization + convolution + ReLU + pooling operation to complete downsampling operation;

(3) performing continuous operation of deconvolution, BN and an activation function on the feature map obtained by the down-sampling to finish the up-sampling;

(4) in the same network layer, the output characteristic mapping of the down sampling and the up sampling are connected; fusing the output characteristic diagrams of down-sampling and up-sampling of the next neural network layer from the top layer neural network layer to the bottom layer neural network layer in different network layers, continuously connecting the fused output characteristic diagrams by the up-sampling characteristic diagram of the next neural network layer, and continuously iterating until the next layer has no corresponding up-sampling, and then obtaining the mapping relation between 89GHz data and 36GHz data with high reliability,

(5) inputting the test set data, namely the relationship between the undisturbed 89GHz data and the 36GHz data, and the mapping relationship between the 89GHz data and the highly reliable 36GHz data obtained in the step (4) into a discrimination network;

(6) then, continuous batch normalization + convolution + ReLU + pooling operation is carried out to complete down-sampling operation;

(7) finally, judging the result obtained by the judgment network by using the cross entropy, outputting the corrected 89GHz data if the loss function reaches the minimum value, otherwise, returning to the step (2), and repeating the steps until the loss function reaches the minimum value;

and S3, inverting the sea ice density by adopting an ASI sea ice density algorithm based on the corrected 89GHz data.

The characteristics of different network layers can be integrated by the Unet network through a jump connection mode, so that the denoising performance is improved, and the Unet network has stronger self-adaptability and can effectively retain structural information of images, so that the Unet network is adopted as a generation network G;

wherein, Unet includes input layer, convolution layer, pooling layer, activation layer and output layer;

specifically setting the kernel of the pooling layer to be 2 × 2, setting the size of the convolution filter to be 3 × 3, and selecting the ReLU as an activation function;

the discrimination network D has the function of discriminating two groups of mapping relations, the discrimination network D adopts a CNN network, the image is subjected to feature extraction through continuous down-sampling layers after being input into the discrimination network, the sizes of convolution kernels in the down-sampling layers are all 4 multiplied by 4, the steps are all 2, in order to prevent the situation that the gradient disappears when the parameters are updated by using a gradient descent method, the convolution operation of the down-sampling layers is subjected to normalization processing, nonlinear mapping is carried out by using a ReLU activation function, finally, the value of each pixel is output by using a Sigmoid function in the discrimination layer, and the final loss is obtained by using cross entropy calculation.

Specifically, the Antarctic sea ice density is inverted to 98 by an ASI algorithm by using AMSR-2 brightness temperature data corrected at 2 months and 1 day in 2021.

FIG. 3 shows the results of sea ice density obtained by the method of the present invention.

FIG. 5 is a graph of sea ice concentration results from the method of the present invention for selected areas of FIG. 3.

Comparative example 1

And (3) inverting the south pole sea ice intensity by using uncorrected AMSR-2 bright temperature data of 2 months and 1 day in 2021 through an ASI algorithm, namely the ASI algorithm and a weather filter.

FIG. 4 shows the results of the sea ice concentration obtained by the method of this comparative example.

FIG. 6 is a graph of sea ice concentration results from the comparative example method for selected regions of FIG. 4.

Comparative example 2

And (3) inverting the south pole sea ice density by using uncorrected AMSR-2 brightness temperature data of 2 months and 1 days in 2021 through an ASI algorithm, and inverting the south pole sea ice density to 88 through the ASI algorithm.

Comparative example 3:

and (5) verifying the result by using high-resolution optical remote sensing data Landsat8 OLI.

Based on Landsat8 data (resolution: 30m), 160-175E and 74-79S near Ross sea were selected for further validation. Based on Landsat8 data, we used NSDI (normalized Difference snow index) calculated by green and short band infrared band thresholding to detect sea ice distribution, which can achieve ice water identification based on differences in the reflectivity of sea ice and sea water in the red and near infrared regions (Perovich, 1996; Riggs et al, 1999; Riggs and Hall, 2015; Hall et al, 2001; Liu et al, 2016). Then, the proportion of the number of sea ice pixels in the corresponding AMSR-2 pixel grid is counted, and the proportion is used as the sea ice density result of Landsat8 OLI, and is specifically 99.

The results of the sea ice distribution obtained from Landsat8 OLI data using reflectance thresholding are shown in fig. 7. In fig. 7, the white area is sea ice, the black area is sea water, and the gray area is land.

And (3) verification and analysis:

the results of the example and comparative examples 1-2 show that the sea ice density results of the ASI algorithm inversion based on the calibration data are closer to the sea ice density results obtained by the ASI algorithm plus the weather filter. To further validate the sea ice concentration results of the ASI algorithm inversion based on the corrected data, we validated the sea ice concentration results obtained using the high resolution Landsat8 data (resolution 30 m). The sea ice concentration calculated using the ASI algorithm based on the calibration data is 98 which is closer to the sea ice concentration 99 obtained from the high resolution data than the sea ice concentration 88 calculated using the ASI algorithm.

The ASI algorithm based on the correction data significantly changes the sea ice concentration of the mixed pixels, thereby reducing the weather effect on the high frequency data. Compared with the sea ice density obtained by the ASI algorithm and the weather filter, the sea ice density result inverted by the ASI algorithm based on the data after the GAN correction has higher precision.

The sea ice data correction method can also be applied to other data sources, and provides new method support for sea ice data correction based on microwave radiometer data.

While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (3)

1. The high-precision ASI sea ice density inversion algorithm for data correction based on CGAN is characterized by comprising the following steps of: the method comprises the following steps:

s1, data screening: finding a relatively stable relation between 89GHz data and 36GHz data which are not interfered by external environments such as cloud, water vapor and the like or are slightly interfered, and screening out 89GHz data which are greatly interfered;

s2, data correction: the generation network and the judgment network of the CGAN correct 89GHz data with large interference by using a stable relation between 36GHz data with high reliability and 89GHz data which is not interfered by external environments such as cloud, water vapor and the like or has small interference through countermeasure training;

the model function of CGAN is shown in (3):

wherein, x is data influenced by external environment, y is additional information, z is input random noise, G (z | y) is data which is not influenced by external environment and is input to the generation network of the CGAN under the influence of external environment, and D (G (z | y)) is the probability of judging the input data to be false by the judgment network; since the goal of the CGAN's generation network is to make the generated data as close as possible to data that is not affected by the external environment, the penalty function is set to 1-D (G (z | y)) to ensure that the discrimination network outputs false image probabilities as small as possible, and the goal of the discrimination network is to improve the ability to discriminate differences in the input data, so the larger D (x | y) the better, while the smaller the noise impact is desired to be the better, the penalty function is set to D (x | y) +1-D (G (z | y)), and so on

Indicating the progress of the game;

the CGAN model is trained as follows:

(1) before training, adjusting the data sets, such as rotating and translating, increasing the number of the data sets, and then carrying out normalization processing on the training set and the test set;

(2) inputting the training set into a generating network, and then carrying out continuous batch normalization + convolution + ReLU + pooling operation to complete downsampling operation;

(3) performing continuous operation of deconvolution, BN and an activation function on the feature map obtained by the down-sampling to finish the up-sampling;

(4) in the same network layer, the output characteristic mapping of the down sampling and the up sampling are connected; fusing the output characteristic diagrams of down-sampling and up-sampling of the next neural network layer from the top layer neural network layer to the bottom layer neural network layer in different network layers, continuously connecting the fused output characteristic diagrams by the up-sampling characteristic diagram of the next neural network layer, and continuously iterating until the next layer has no corresponding up-sampling, and then obtaining the mapping relation between 89GHz data and 36GHz data with high reliability,

(5) inputting the test set data, namely the relationship between the undisturbed 89GHz data and the 36GHz data, and the mapping relationship between the 89GHz data and the highly reliable 36GHz data obtained in the step (4) into a discrimination network;

(6) then, continuous batch normalization + convolution + ReLU + pooling operation is carried out to complete down-sampling operation;

(7) finally, judging the result obtained by the judgment network by using the cross entropy, outputting the corrected 89GHz data if the loss function reaches the minimum value, otherwise, returning to the step (2), and repeating the steps until the loss function reaches the minimum value;

and S3, inverting the sea ice density by adopting an ASI sea ice density algorithm based on the corrected 89GHz data.

2. The CGAN-based high accuracy ASI sea ice density inversion algorithm of claim 1 further comprising: the method for screening 89GHz data with large interference is specifically,

(1) under the condition of clear weather, a polarization ratio scatter diagram is drawn by taking a 36GHz data polarization ratio as an abscissa and an 89GHz data polarization ratio as an ordinate;

wherein the polarization ratio formula is shown as (1),

wherein TBV and TBH are respectively vertical polarization bright temperature and horizontal polarization bright temperature;

(2) in the polarization ratio scattergram, the abscissa is equally divided into several intervals, and the mean value and standard deviation of the ordinate in each interval are calculated, and then, a least squares best fit curve is drawn using the mean value minus two times the standard deviation in each interval, the curve being approximated by a quadratic equation, as shown in (2),

PR₈₉＝a(PR₃₆)²-bPR₃₆+c (2)

wherein PR89 is the polarization ratio of 89GHz data, PR36 is the polarization ratio of 36GHz data, and a, b and c are constants.

3. The CGAN-based high accuracy ASI sea ice density inversion algorithm of claim 1 further comprising: the characteristics of different network layers can be integrated by the Unet network in a jumping connection mode, so that the denoising performance is improved, and the Unet network has stronger self-adaptability and can effectively retain structural information of images, so that the Unet network is adopted as a generation network G;

wherein, Unet includes input layer, convolution layer, pooling layer, activation layer and output layer;

specifically setting the kernel of the pooling layer to be 2 × 2, setting the size of the convolution filter to be 3 × 3, and selecting the ReLU as an activation function;

the discrimination network D has the function of discriminating two groups of mapping relations, the discrimination network D adopts a CNN network, the image is subjected to feature extraction through continuous down-sampling layers after being input into the discrimination network, the sizes of convolution kernels in the down-sampling layers are all 4 multiplied by 4, the steps are all 2, in order to prevent the situation that the gradient disappears when the parameters are updated by using a gradient descent method, the convolution operation of the down-sampling layers is subjected to normalization processing, nonlinear mapping is carried out by using a ReLU activation function, finally, the value of each pixel is output by using a Sigmoid function in the discrimination layer, and the final loss is obtained by using cross entropy calculation.

Images