Image tone mapping noise suppression method

Abstract

The invention provides an image tone mapping noise suppression method. The method comprises the following steps: S1, acquiring noise modeling data; s2, calibrating noise parameters; s3, preparing tone mapping data; s4, designing a neural network; s5, training strategies are involved; and S6, training the model. According to an existing noise modeling method, a tone mapping task and a denoising task are combined together, and dynamic range compression and noise suppression are completed when a neural network is used for tone mapping, so that residual noise and smear traces are not obviously amplified when a high dynamic range image is converted into a low dynamic range image. Noise reduction is facilitated, and the efficiency of the whole intelligent monitoring video processing application system is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring video processing, and particularly relates to an image tone mapping noise suppression method. Background Art

[0002] With the development of technology, especially the neural networks applicable to fields such as image recognition, face recognition, video surveillance processing, etc. have been increasingly widely used with the breakthrough development of artificial intelligence. High-dynamic range images can reflect real scenes more meticulously, but high-dynamic range images are larger than low-dynamic range images, requiring larger storage space and transmission bandwidth, and currently most graphics output devices such as monitors and printers have a smaller dynamic range, making it difficult to output and display high-dynamic range images. Therefore, tone mapping is needed for conversion.

[0003] During the process of an image sensor converting photons into the gray values of digital images, various noises will be generated, including photon shot noise, row noise, readout noise, etc. Currently, there are existing methods to model such noises.

[0004] Currently, there are some tone mapping methods such as traditional curve mapping and neural network model mapping to convert high-dynamic range images into low-dynamic range images.

[0005] However, if noise suppression or denoising is not performed, tone mapping will amplify the noise as a signal, resulting in a poor visual effect. Therefore, the current mainstream method is to first denoise the image and then perform tone mapping; but if the degree of noise removal is small, there will be more residual noise, and if the degree of removal is large, details will be lost, resulting in smear marks, and general tone mapping will amplify the residual noise and smear marks.

[0006] In addition, the commonly used technical terms in the prior art include:

[0007] Tone mapping: Tone mapping enables high-dynamic range images to be displayed on our low-dynamic range monitors and conforms to the human visual experience as much as possible. Its essence is to convert high-dynamic range images into low-dynamic range images.

[0008] Noise: When an image sensor works, a certain number of photons incident on the pixel surface during the exposure time are converted into a certain number of electrons, which are converted into a voltage signal of a certain amplitude through a capacitor, amplified and quantized, and finally become the gray values of digital images, during which various noises will be generated;

[0009] Tone mapping noise suppression: When converting high-dynamic range to low-dynamic range in the tone mapping task, if noise suppression is not performed, the noise will be amplified as a signal, resulting in a poor visual effect. Summary of the Invention

[0010] To solve the above problems, the purpose of this application is: according to the existing noise modeling method, enable the neural network to learn the noise form, reduce the residual noise and smear marks caused by denoising, and at the same time complete the tone mapping task.

[0011] Specifically, the present invention provides an image tone mapping noise suppression method, and the method includes:

[0012] S1. Collect data required for noise modeling, including black frame data and flat frame data:

[0013] S1.1. Collect black frame data, that is, completely cover the lens with black tape, and collect 20 raw images respectively at isos of 100, 200, 400, 800, 1600, and 3200.

[0014] S1.2. Collect flat frame data, that is, fix the image sensor without moving, face the lens directly at the uniform white paper under uniform ambient light, and at isos of 100, 200, 400, 800, 1600, and 3200 respectively, with 1 as the starting point of the exposure row, and the non-overexposed exposure row, that is, the exposure row with the maximum value of the 12-bit raw image data not exceeding 4095 as the end point, and sample 20 groups of data at an average interval, and each group of data includes 2 raw images.

[0015] S2. Calibrate noise modeling parameters:

[0016] S2.1. Calibrate the system gain k parameter, select the 256x256 area of each flat frame collected in step S1.2 and denote it as I f1 , I f2 , and calculate the mean m and variance v according to formula (1), where Mean represents calculating the mean and Var represents calculating the variance.

[0017]

[0018] For each iso, perform formula (1) on 20 groups of data respectively to obtain 20 groups of parameters, and perform linear regression with m and v as the horizontal and vertical axes, and the obtained slope is the system gain k under the current iso.

[0019] S2.2. Calibrate the row noise N row parameter, select the black frame collected in step S1.1 under the same iso, as shown in formula (2), subtract the black level from the black frame and use the standard normal distribution for fitting to obtain σ n , where I d is the black frame, bl is the black level, and fit n is the normal distribution fitting function, and the corresponding σ can be obtained for each iso n , and use the iso corresponding to log(k), log(σ n)The slope a is obtained by performing linear regression on the horizontal axis x and the vertical axis y n and the intercept b n , and the linear regression is shown in Equation (3);

[0020] σ n = fit n (I d - bl) Equation (2)

[0021] y = a n x + b n Equation (3);

[0022] S2.3, calibrate the read noise N read parameters. Under the same iso, select the black frame collected in step S1.1, subtract the black level and the row noise that follows the standard normal distribution, and then use the tukeylambda distribution for fitting to obtain tlshape and σ tl , as shown in Equation (4), where r n represents the row noise that follows the standard normal distribution, and fit tl is the tukeylambda distribution fitting function. Corresponding tlshape and σ can be obtained for each iso tl , and with iso corresponding to log(k), log(σ tl ), the slope a is obtained by performing linear regression on the horizontal axis x and the vertical axis y according to Equation (3) tl and the intercept b tl ;

[0023]

[0024] S3. Prepare tone mapping data, which consists of 600 pairs of high-dynamic range images and low-dynamic range images to form training pairs, where the training data are all clean and noise-free data;

[0025] S4. Design the neural network structure:

[0026] S4.1. Design the basic feature extraction module. Use 1 convolution with a kernel size of 3x3, a stride of 1, and a padding of 1, and 2 convolutions with a kernel size of 3x3, a stride of 2, and a padding of 1 in series for downsampling. After each convolution, a relu activation function is connected. The original image passes through the basic feature extraction module to obtain basic features;

[0027] S4.2. Design the global feature extraction module. Use 3 consecutive convolutions with a kernel size of 3x3, a stride of 2, and a padding of 1 in series. After each convolution, a relu activation function is connected. After the basic features pass through the global feature module, the mean value is finally calculated in the width w and height h directions to obtain the global features;

[0028] S4.3. Design a scaling and offset module, which consists of 6 fully connected layers. The global features generated in step S4.2 are input into the first 2 fully connected layers after being convolved by a convolutional layer with a kernel size of 3x3, a stride of 1, and a padding of 1, generating a set of scaling parameters and offset parameters. According to Equation (5), the features are mapped using the scaling and offset parameters, where f in is the input feature, scale and shift are the scaling and offset parameters respectively, and f out is the output feature, and then upsampled by nearest neighbor interpolation; similarly, through the middle 2 fully connected layers and nearest neighbor interpolation upsampling; finally, the full-resolution features are obtained after passing through the last 2 fully connected layers;

[0029] f out = f in ·scale + shift + f in Equation (5)

[0030] S4.4. Design a compression parameter module. The base features are upsampled by nearest neighbor interpolation to obtain full-resolution features. These features are concatenated and fused with the features of the scaling and offset module, then convolved by a convolutional layer with a kernel size of 3x3, a stride of 2, and a padding of 1, and finally passed through the tanh function to obtain the full-image dynamic range compression parameter c;

[0031] S4.5. As shown in Equation (6), the full-image dynamic range compression parameter c and the original image I o are multiplied point by point to obtain the final result I f ;

[0032] I f = c·I o Equation (5)

[0033] S5. Design a training strategy:

[0034] S5.1. Read the raw data and randomly select to generate a noise region mask roi ;

[0035] S5.2. According to Equation (6), determine the uniform distribution random sampling range based on the system gain k of different ISOs. Equation (6) is as follows, where log(k s ) is the selected system gain parameter, U is the uniform distribution sampling, and k min and k max are the minimum and maximum system gains under different ISOs;

[0036] log(k s ) ~ U(log(k min ), log(k max )) Equation (6)

[0037] S5.3, Generate random noise N on the input high-dynamic range data I using log(k s ) as the Poisson distribution parameter. This noise belongs to the generated photon shot noise; HDR p

[0038] S5.4, Sample the readout noise parameter from a normal distribution according to Equation (7) as follows. Here, log(σ ns ) is the selected row noise parameter. Use σ n to generate random noise N through a standard normal distribution. This noise belongs to the row noise; row

[0039] log(σ ns ) ~ N(a n ·log(k s ) + b n , σ n ) Equation (7)

[0040] S5.5, Sample the readout noise parameter from a normal distribution according to Equation (8) as follows. Here, log(σ tls ) is the selected readout noise parameter. Use tlshape, σ tls to generate random noise N through a Tukey lambda distribution. This noise belongs to the readout noise; read

[0041] log(σ tls ) ~ N(a tl ·log(k s ) + b tl , σ tl ) Equation (8)

[0042] S5.6, Use Equation (9) to add up the noises generated in the above steps to obtain the total noise.

[0043] N = N p + N row + N read Equation (9)

[0044] S5.7, Add the same noise to the random noise regions of the tone mapping data training pairs respectively, as shown in Equation (10). Here, I LDR , I HDR are the low-dynamic range data and high-dynamic range data of the tone mapping respectively. I' LDR , I' HDR are the low-dynamic range data and high-dynamic range data after adding the same noise respectively

[0045] ​​​​

[0046] S6. Train the model according to the neural network designed in step S4 and the training strategy designed in step S5.

[0047] In step S6, the optimizer uses Adam, the learning rate is 0.0001, the training cycle is 500, the learning rate is reduced by 0.1 times every 200 cycles, and the training loss is L1 loss.

[0048] The image sensor is imx327.

[0049] The black level of the imx327 image sensor is 240.

[0050] Therefore, the advantages of this application are as follows: According to the existing noise modeling method, the tone mapping task and the denoising task are combined. When using a neural network for tone mapping, dynamic range compression and noise suppression are completed, so that when a high dynamic range image is converted into a low dynamic range image, the residual noise and smear marks are not significantly amplified. It is beneficial to noise reduction and improves the efficiency of the entire intelligent surveillance video processing application system. Brief Description of the Drawings

[0051] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention.

[0052] Figure 1 It is a schematic structural diagram of the system of this application.

[0053] Figure 2 It is a schematic diagram of the neural network structure involved in this application. Detailed Embodiment

[0054] In order to more clearly understand the technical content and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings.

[0055] This application proposes an image tone mapping noise suppression method. As Figure 1 shown, the image sensor used in this method is imx327, and the main implementation steps of the method are as follows:

[0056] Step S1, collect the data required for noise modeling, including black frame data and flat frame data:

[0057] S1.1, collect black frame data. Completely cover the lens with black tape and collect 20 raw images respectively at isos of 100, 200, 400, 800, 1600, and 3200.

[0058] S1.2, Collect flat frame data. That is, fix the image sensor without movement. With the lens facing a uniform white paper under uniform ambient light, at isos of 100, 200, 400, 800, 1600, and 3200 respectively, starting from exposure row 1, using the non-overexposed exposure row, that is, the exposure row where the 12-bit raw image data value does not exceed 4095 at most as the end point, sample 20 groups of data at an average interval, and each group of data includes 2 frames of raw images;

[0059] Step S2, Calibrate the noise modeling parameters:

[0060] S2.1, Calibrate the system gain k parameter. Select the 256x256 area of each flat frame collected in step S1.2 and denote it as I f1 , I f2 , Calculate the mean m and variance v according to formula (1), where Mean represents calculating the mean and Var represents calculating the variance.

[0061]

[0062] For each iso, perform formula (1) on the 20 groups of data respectively to obtain 20 groups of parameters. Use m and v as the horizontal and vertical axes for linear regression, and the obtained slope is the system gain k under the current iso;

[0063] Formula (1) can be represented by the following Python code:

[0064] import numpy as np

[0065] from scipy import stats

[0066] img_sum = img_1 + img_2

[0067] img_diff = img_1 - img_2

[0068] m = np.mean(img_sum)

[0069] v = np.var(img_diff);

[0070] S2.2, Calibrate the row noise N row parameter. Select the black frame collected in step S1.1 under the same iso. As shown in formula (2), after subtracting the black level from the black frame, use the standard normal distribution for fitting to obtain σ n , where I d is the black frame, bl is the black level, and the black level of the imx327 image sensor is 240, and fit n is the normal distribution fitting function. The corresponding σ can be obtained for each iso n, with iso corresponding to log(k), log(σ n ) as the horizontal axis x and the vertical axis y to perform linear regression to obtain the slope a n and the intercept b n , the linear regression is shown in Equation (3);

[0071] σ n = fit n (I d - bl) Equation (2)

[0072] y = a n x + b n Equation (3);

[0073] fit n can be represented by the following Python code:

[0074] from scipy.stats import norm

[0075] σ n = norm.fit(img);

[0076] S2.3, calibrate the read noise N read Parameter, under the same iso, select the black frame collected in step S1.1, subtract the black level and the row noise that follows the standard normal distribution, and then use the tukeylambda distribution for fitting to obtain tlshape and σ tl , as shown in Equation (4), where r n represents the row noise that follows the standard normal distribution, fit tl is the tukeylambda distribution fitting function, and corresponding tlshape and σ can be obtained for each iso tl , with iso corresponding to log(k), log(σ tl ) as the horizontal axis x and the vertical axis y, perform linear regression according to Equation (3) to obtain the slope a tl and the intercept b tl ;

[0077]

[0078] fit tl can be represented by the following Python code:

[0079] from scipy import stats

[0080] res_max = stats.ppcc_max(img, dist='tukeylambda')

[0081] tlshape and σtl = stats.probplot(img, sparams=(res_max,), dist='tukeylambda',

[0082] fit=True, plot=None, rvalue=False);

[0083] Step S3, prepare tone mapping data, including 600 pairs of high-dynamic range images and low-dynamic range images to form training pairs, where the training data are all clean and noise-free data; the method of this application lies in noise suppression in tone mapping, so the data acquisition is achieved on the basis of existing tone mapping data, which will not be elaborated here.

[0084] Step S4, design the neural network structure, as Figure 2 shown, including:

[0085] S4, design the neural network structure:

[0086] S4.1, design the basic feature extraction module, use 1 convolution with a kernel size of 3x3, stride of 1, and padding of 1, and 2 convolutions with a kernel size of 3x3, stride of 2, and padding of 1 in series for downsampling. After each convolution, a relu activation function is connected. The original image passes through the basic feature extraction module to obtain basic features;

[0087] S4.2, design the global feature extraction module, use 3 consecutive convolutions with a kernel size of 3x3, stride of 2, and padding of 1 in series. After each convolution, a relu activation function is connected. The basic features pass through the global feature module and finally obtain the global features by taking the mean in the width w and height h directions. Taking the mean is a basic operation and will not be elaborated as an existing technology;

[0088] S4.3, design the scaling and offset module, use 6 fully connected layers to form the scaling and offset module. The global features generated in step S4.2 pass through 1 convolution with a kernel size of 3x3, stride of 1, and padding of 1 and then are input into the first 2 fully connected layers to generate 1 set of scaling parameters and offset parameters. According to Equation (5), use the scaling and offset parameters to map the features, where, f in is the input feature, scale and shift are the scaling and offset parameters respectively, f out is the output feature, and then perform upsampling through nearest neighbor interpolation. Nearest neighbor interpolation is a basic operation; similarly, pass through the middle 2 fully connected layers and nearest neighbor interpolation upsampling; finally, obtain the full-resolution features after passing through the last 2 fully connected layers;

[0089] f out = f in ·scale + shift + fin Equation (5)

[0090] S4.4. Design the compression parameter module. The basic features are upsampled by nearest neighbor interpolation to obtain full-resolution features. These features are concatenated (a basic neural network operator) with the features of the scaling and offset module and then fused. After that, they are convolved with a convolution kernel of 3x3, a stride of 2, and a pad of 1, and then passed through the tanh function (a basic neural network operator, no formula required) to obtain the full-image dynamic range compression parameter c.

[0091] S4.5. As shown in Equation (6), the full-image dynamic range compression parameter c and the original image I o are multiplied point by point to obtain the final result I f ;

[0092] I f = c · I o Equation (5)

[0093] S5. Design the training strategy:

[0094] S5.1. Read the raw data and randomly select to generate a noise region mask roi ;

[0095] S5.2. Determine the uniform distribution random sampling range according to Equation (6) with different system gain k values for different isos. Equation (6) is as follows, where log(k s ) is the selected system gain parameter, U is the uniform distribution sampling, and k min and k max are the minimum and maximum system gain values for different isos;

[0096] log(k s ) ~ U(log(k min ), log(k max )) Equation (6)

[0097] S5.3. Use log(k s ) as the Poisson distribution parameter to generate random noise N HDR on the input high-dynamic range data I p , and this noise belongs to the generated photon shot noise;

[0098] S5.4. Obtain the readout noise parameter by sampling from a normal distribution according to Equation (7). Equation (7) is as follows, where log(σ ns ) is the selected row noise parameter, and σ n is used to generate random noise N row through the standard normal distribution, and this noise belongs to the row noise;

[0099] log(σ ns ) ~ N(an · log(k s ) + b n , σ n ) Equation (7)

[0100] S5.5, Sample the read noise parameter according to the normal distribution using Equation (8). Equation (8) is as follows, where log(σ tls ) is the selected read noise parameter, use tlshape, σ tls Generate random noise N through the Tukey lambda distribution read , and this noise belongs to the read noise

[0101] log(σ tls ) ~ N(a tl · log(k s ) + b tl , σ tl ) Equation (8)

[0102] S5.6, Use Equation (9) to add up the noises generated in the above steps to obtain the total noise

[0103] N = N p + N row + N read Equation (9)

[0104] S5.7, Add the same noise to the random noise regions of the tone mapping data training pairs respectively, as shown in Equation (10), where I LDR , I HDR are the low dynamic range data and high dynamic range data of the tone mapping respectively, and I' LDR , I' HDR are the low dynamic range data and high dynamic range data after adding the same noise respectively

[0105]

[0106] Step S6, Perform model training according to the neural network designed in Step S4 and the training strategy designed in Step S5. Use Adam as the optimizer, the learning rate is 0.0001, the training epoch is 500, the learning rate is reduced by 0.1 times every 200 epochs, and the training loss is L1 loss

[0107] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention

Claims

1. An image tone mapping noise suppression method, characterized in that, the method includes: S1. Collect data required for noise modeling, including black frame data and flat frame data: S1.

1. Collect black frame data, that is, completely cover the lens with black tape, and collect 20 raw images respectively at isos of 100, 200, 400, 800, 1600, and 3200. S1.

2. Collect flat frame data, that is, fix the image sensor without moving, face the lens towards a uniform white paper under uniform ambient light, and respectively at isos of 100, 200, 400, 800, 1600, and 3200, with 1 as the starting point of the exposure row and the non-overexposed exposure row, that is, the exposure row with the maximum value of the 12-bit raw image data not exceeding 4095 as the end point, and sample 20 groups of data at an average interval, with each group of data including 2 raw images. S2. Calibrate noise modeling parameters: S2.1, calibrate the system gain k parameter, and select the 256x256 area of each flat frame collected in step S1.2 and denote it as I f1 , I f2 , calculate the mean m and variance v according to formula (1), where Mean represents calculating the mean and Var represents calculating the variance For each iso, perform equation (1) on 20 groups of data respectively to obtain 20 groups of parameters, and perform linear regression with m and v as the horizontal and vertical axes to obtain the slope, which is the system gain k under the current iso. S2.2, calibrate the row noise N row Parameter, select the black frame collected in step S1.1 under the same iso. As shown in Equation (2), after subtracting the black level from the black frame, σ is obtained by fitting with the standard normal distribution n , where I d is the black frame, bl is the black level, and fit n is the normal distribution fitting function. The corresponding σ can be obtained for each iso n , with iso corresponding to log(k), log(σ n ) as the horizontal axis x and the vertical axis y for linear regression to obtain the slope a n and the intercept b n , and the linear regression is as shown in Equation (3); σ n = fit n (I d - bl) Equation (2) y = a n x + b n Formula (3); S2.3, calibrate the read noise N read Parameter, select the black frame collected in step S1.1 under the same iso, subtract the black level, and use the tukeylambda distribution to fit the row noise that follows the standard normal distribution to obtain tlshape and σ tl , as shown in Equation (4), where r n represents the row noise that follows the standard normal distribution, and fit tl is the tukeylambda distribution fitting function. Corresponding tlshape and σ can be obtained for each iso tl , with iso corresponding to log(k), log(σ tl ) as the horizontal axis x and the vertical axis y, perform linear regression according to Equation (3) to obtain the slope a tl and the intercept b tl ; S3. Prepare tone mapping data, including 600 pairs of high-dynamic-range images and low-dynamic-range images to form training pairs, where the training data are all clean and noise-free data. S4. Design a neural network structure: S4.

1. Design a basic feature extraction module, use 1 convolution with a convolution kernel of 3x3, stride of 1, and pad of 1 and 2 convolutions with convolution kernels of 3x3, stride of 2, and pad of 1 in series for downsampling, and connect a relu activation function after each convolution. The original image passes through the basic feature extraction module to obtain basic features. S4.

2. Design a global feature extraction module, use 3 consecutive convolutions with convolution kernels of 3x3, stride of 2, and pad of 1 in series, connect a relu activation function after each convolution, and finally calculate the mean in the width w and height h directions after the basic features pass through the global feature module to obtain the global features. S4.3, Design a scaling and offset module, which consists of 6 fully connected layers. The global features generated in step S4.2 are input into the first 2 fully connected layers after being convolved by a convolutional kernel of 3x3, stride of 1, and pad of 1, generating a set of scaling parameters and offset parameters. According to Equation (5), the features are mapped using the scaling and offset parameters, where f in is the input feature, scale and shift are the scaling and offset parameters respectively, and f out is the output feature, and then upsampling is performed by nearest neighbor interpolation; similarly, through the middle 2 fully connected layers and nearest neighbor interpolation upsampling; finally, the full-resolution features are obtained after passing through the last 2 fully connected layers; f out = f in · scale + shift + f in Equation (5) S4.

4. Design a compression parameter module. The basic features are upsampled by nearest neighbor interpolation to obtain full-resolution features. These features are concatenated and fused with the features of the scaling offset module and then pass through a convolution with a convolution kernel of 3x3, stride of 2, and pad of 1, and then pass through the tanh function to obtain the full-image dynamic range compression parameter c. S4.5, as shown in Equation (6), the full-image dynamic range compression parameter c and the original image I o are multiplied point by point to obtain the final result I f ; I f = c·I o Equation (5) S5. Design a training strategy: S5.1, Read the raw data and randomly select to generate a noise region mask roi ; S5.

2. Determine the uniformly distributed random sampling range according to the system gain k of different ISOs according to Equation (6). Equation (6) is as follows, where log(k s ) is the selected system gain parameter, U is uniformly distributed sampling, k min and k max are the minimum and maximum values of the system gain under different ISOs; log(k s )~U(log(k min ),log(k max )) Equation (6) S5.3, using log(k s ) as the Poisson distribution parameter to generate random noise N HDR on the input high-dynamic-range data I p , where the noise belongs to the generated photon shot noise; S5.4, sample the readout noise parameter according to the normal distribution in Equation (7), Equation (7) is as follows, where log(σ ns ) is the selected row noise parameter, and σ n is used to generate random noise N row through the standard normal distribution, and this noise belongs to the row noise; log(σ ns ) ~ N(a n ·log(k s ) + b n , σ n ) Equation (7) S5.5, sample the readout noise parameter according to the normal distribution using Equation (8) as follows, where log(σ tls ) is the selected readout noise parameter, using tlshape, σ tls Generate random noise N through the Tukey lambda distribution read , which belongs to the readout noise log(σ tls ) ~ N(a tl ·log(k s ) + b tl , σ tl ) Equation (8) S5.

6. Use equation (9) to add up the noises generated in the above steps to obtain the total noise. N = N p + N row + N read Formula (9) S5.7, Add the same noise to the random noise regions of the tone mapping data training pairs respectively, as shown in Equation (10), where I LDR , I HDR are the low-dynamic range data and high-dynamic range data of the tone mapping respectively, and I' LDR , I' HDR are the low-dynamic range data and high-dynamic range data after adding the same noise respectively; S6. Perform model training according to the neural network designed in step S4 and the training strategy designed in step S5.

2. The image tone mapping noise suppression method according to claim 1, characterized in that, in step S6, the optimizer uses Adam, the learning rate is 0.0001, the training cycle is 500, the learning rate is reduced by 0.1 times every 200 cycles, and the training loss is L1 loss.

3. The image tone mapping noise suppression method according to claim 1, characterized in that, the image sensor is imx327.

4. A method for suppressing noise in image tone mapping according to claim 3, characterized in that, the black level of the imx327 image sensor is 240.

Images