A method for modeling discrete noise of dark light images

Abstract

The application provides a modeling method of discrete noise points of a dark light image, comprising the following steps: S1: using a target camera to collect multiple black frames, and performing black level correction on the black frames; S2: extracting discrete noise points from the corrected black frames based on criteria, wherein the criteria include but are not limited to 3sigma criteria and 6sigma criteria; S3: calculating a probability mass function (PMF) and a cumulative distribution function (CDF) of each channel according to the extracted discrete noise points; S4: generating simulated discrete noise points through inverse transform sampling based on the CDF; and S5: outputting: generating an image with simulated discrete noise points. The discrete noise points beyond the Gaussian distribution representation range are modeled separately, so that the noise modeling under the dark light condition is more accurate, the deviation between the synthesized data and the real noise distribution is reduced, and the denoising effect of the denoising model in the dark light environment is improved.

Description

Technical Field

[0001] This invention belongs to the field of video noise reduction technology, and specifically relates to a method for modeling discrete noise points in low-light images. Background Technology

[0002] With the rapid development of digital imaging technology, the demand for image acquisition in low-light environments is increasing. However, under low-light conditions, image sensors exhibit large-scale noise (i.e., discrete noise) with complex distribution characteristics in images due to various factors such as raw materials and manufacturing processes. This noise not only exhibits non-uniform spatial distribution characteristics but also couples with various noise sources such as sensor thermal noise and readout circuit noise, posing a severe challenge to denoising methods.

[0003] Current mainstream denoising techniques can be broadly categorized into two types: traditional algorithms and deep learning methods. Traditional algorithms, such as BM3D and nonlocal mean denoising, are based on statistical priors and exhibit good denoising performance under normal lighting conditions. However, they have significant shortcomings when processing low-light images: First, the amplitude distribution of discrete noise points often exceeds the representation range of the traditional Gaussian-Poisson noise model; second, the high degree of aliasing between noise and signal in the frequency domain leads to severe loss of edge details; and third, the algorithm parameters require manual adjustment, making it difficult to adapt to the noise characteristics of different imaging devices.

[0004] While deep learning-based denoising methods (such as DnCNN and CBDNet) exhibit stronger feature learning capabilities, their performance is highly dependent on the quality of the training data. Existing methods generally employ data synthesis based on simple noise models, which cannot accurately simulate the nonlinear superposition characteristics of discrete noise points in low-light environments. This leads to significant shortcomings in practical applications of the trained convolutional neural networks: insufficient suppression of strong noise regions, susceptibility to artifacts, incomplete recovery of high-frequency details, and poor generalization ability to device-specific noise.

[0005] However, the existing technology has the following drawbacks:

[0006] In terms of noise modeling, existing technologies still have limitations: they either only consider the linear superposition of a single noise source or use empirical noise parameter estimation methods, failing to establish a discrete noise generation model that conforms to the physical imaging process. This directly restricts the practical application effect of deep learning denoising models in low-light scenes.

[0007] In addition, the technical terms in this field include:

[0008] RAW image: refers to the raw image data output by the image sensor without any processing or decoding, existing in the form of uncompressed, raw pixel information.

[0009] Black frame: RAW image captured by the camera in a dark environment.

[0010] PMF: Probability Mass Function (PMF) is a function that describes the probability of a discrete random variable for each specific value.

[0011] CDF: The Cumulative Distribution Function (CDF) is a function that describes the probability that a random variable X takes a value less than or equal to a specific value x.

[0012] The 3σ criterion is an important rule in statistics for describing the dispersion of normally distributed (or approximately normally distributed) data. For a normal distribution, almost all (about 99.7%) values ​​of the random variable X fall within three standard deviations of the mean. Summary of the Invention

[0013] To address the aforementioned issues, the purpose of this application is to provide a method for modeling discrete noise points in low-light images, based on inverse transform sampling. This method aims to solve the problems in existing technologies where the modeling of discrete noise points in low-light images is inaccurate, and there is a deviation between the synthesized data and the actual noise distribution, resulting in insufficient suppression of strong noise by deep learning denoising networks.

[0014] Specifically, the present invention provides a method for modeling discrete noise points in a low-light image, the method comprising the following steps:

[0015] S1: Acquire multiple black frames using the target camera, and perform black level correction on the black frames; including:

[0016] Multiple RAW format black frames were captured from the target camera in a dark environment.

[0017] Black level correction is performed on black frames, including converting the RAW image to a multi-channel format and subtracting the corresponding black level value from each channel independently.

[0018] S2: Extract discrete noise from the corrected black frame based on criteria, including but not limited to the 3σ and 6σ criteria, including:

[0019] Calculate the mean μ and standard deviation σ of the pixel values ​​for each channel;

[0020] Pixels with values ​​exceeding the range of μ±3σ are classified as discrete noise.

[0021] S3: Based on the extracted discrete noise points, calculate the probability mass function (PMF) and cumulative distribution function (CDF) for each channel; including:

[0022] The frequency of discrete noise points is statistically analyzed and normalized to obtain the PMF;

[0023] CDF is calculated based on PMF;

[0024] S4: Based on CDF, simulated discrete noise is generated through inverse transform sampling; including:

[0025] The number N of simulated discrete noise points is determined based on actual needs. R ;

[0026] N is generated in a uniform distribution in [0,1]. R A random number U R ;

[0027] The amplitude of simulated discrete noise is generated by mapping random numbers using the inverse function of CDF.

[0028] To assign random spatial locations to simulate discrete noise points;

[0029] Generate simulated discrete noise points for R / Gr / Gb / B channels following the steps described above;

[0030] S5: Output: Generates an image with simulated discrete noise.

[0031] Step S1: Acquire multiple black frames using the target camera and perform black level correction on the black frames; the black frames are RAW images acquired by the camera in a dark environment, containing all signal-independent noise such as readout noise and stripe noise. Based on the black frames, the distribution and amplitude of discrete noise points can be specifically analyzed; including:

[0032] S1.1: Acquire multiple RAW format black frames from the target camera in a dark environment. The black frames are acquired by covering the camera lens and saving the RAW image output by the camera in a completely dark room. To avoid random errors, multiple black frames need to be acquired. The number of black frames can be set according to actual conditions. For example, the number of black frames is set to 10.

[0033] S1.2: Perform black level correction on black frames, including converting the RAW image to a multi-channel format and subtracting the corresponding black level value from each channel independently;

[0034] Black frames are corrected for black levels to eliminate the reference signal offset generated by the image sensor under no-light conditions and restore the true image. Considering that the dark current of different color channels, including R / Gr / Gb / B, may be inconsistent, each channel is adjusted independently during correction.

[0035] Assuming the Bayer format of the RAW image is RGGB, first convert the single-channel Bayer format RAW image into a four-channel image, and subtract the corresponding black level from each channel to obtain the black level corrected image;

[0036] XR =X R -BL R (1)

[0037] X Gr =X Gr -BL Gr (2)

[0038] X Gb =X Gb -BL Gb (3)

[0039] X B =X B -BL B (4).

[0040] In step S2, discrete noise points are extracted based on the black frame. The readout noise and stripe noise in the black frame basically follow a standard Gaussian distribution. The amplitude of discrete noise points often exceeds the characterization range of the Gaussian distribution. Therefore, discrete noise points can be regarded as outliers in the black frame.

[0041] Assuming the R channel is a black frame, use X R ={x1,x2,...,x n} represents all pixels in the R channel of the black frame, and n is the total number of pixels in the R channel. Assuming that the 3σ criterion in statistics is used to extract outliers in the black frame, the specific steps are as follows:

[0042] Step S2.1, calculate X R mean μ R and standard deviation σ R ;

[0043]

[0044] Step S2.2, assuming X is extracted based on the 3σ criterion R outlier X o As discrete noise, the criteria for extracting discrete noise can be determined according to the actual situation;

[0045] X R_o =(X R >μ R +3σ R )∪(X R <μ R -3σ R (7)

[0046] Step S2.3: Extract the discrete noise points of the Gr / Gb / B channels according to steps S2.1 and S2.2 to obtain X. Gr_o X Gb_o and X G_o ;

[0047] X Gr_o =(X Gr >μ Gr +3σ Gr )∪(X Gr <μ Gr -3σ Gr (8)

[0048] X Gb_o =(X Gb >μ Gb +3σ Gb )∪(X Gb <μ Gb -3σ Gb (9)

[0049] X G_o =(X G >μ G +3σ G )∪(X G <μ G -3σ G (10).

[0050] In step S3, the discrete noise point X extracted from the R / Gr / Gb / B channels is... R_o X Gr_o X Gb_o and X G_o Calculate the probability mass function and cumulative distribution function for each channel;

[0051] Step S3.1: The distribution of discrete noise points is discrete. The probability mass function (PMF) can be obtained by calculating the frequency histogram. The calculation steps are as follows:

[0052] 1) Statistics X R_o ={x1,x2,...,x k The number of times each value appears in the expression {m1, m2, ..., m} k};

[0053] 2) Normalize the frequency to obtain the probability mass function P(X=x) of the discrete noise point. k );

[0054]

[0055] 3) Calculate the probability mass function of the Gr / Gb / B channel according to steps 1) and 2);

[0056] Step S3.2: Calculate the corresponding cumulative distribution function (CDF) based on the probability mass function of each channel; the specific steps are as follows:

[0057] 1) Discrete noise point X R_o All possible values ​​are arranged in ascending order: x a <x b <... <x n ;

[0058] 2) For each value x i Calculate F(x) i ) = P R (X≤x i That is, accumulating from x a To x i All probability mass function values;

[0059] 3) After calculating all values, the cumulative distribution function F is obtained. R (x)=P R (X≤x);

[0060] Calculate the cumulative distribution function of the Gr / Gb / B channels according to steps 1), 2), and 3).

[0061] Step S4 involves generating simulated discrete noise points using an inverse transform sampling method based on the cumulative distribution function. The specific steps are as follows:

[0062] Step S4.1: Calculate the number of simulated discrete noise points; assuming it's the R channel, the total number of pixels in the R channel is M. R Discrete noise X R_o The total number of pixels is K R The number of discrete noise points N to be generated R The calculation method is as follows:

[0063]

[0064] In equation (12), H and W represent the width and height of the simulated noise image, respectively, and α represents the coefficient that controls the number of discrete noise points. The simulated noise image refers to a noisy image obtained by adding simulated noise to a clean image, and then forming an image pair with the clean image for training a denoising neural network.

[0065] Step S4.2, generate N R random numbers

[0066] Step S4.3, for each random number u i Search for P R (X≤x i )≥u i The smallest x i The value is N, obtained after searching for all random numbers. R A simulated discrete noise point;

[0067] Step S4.4: The location of the discrete noise points in the image is random. After generating the discrete noise points, the number is N. R The simulated discrete noise points are assigned random spatial locations, and the specific steps are as follows:

[0068] 1) Randomly select N from the range of 0 to H*W R 1 integer as the position index;

[0069] 2) N R The discrete noise points of the simulation are randomly shuffled;

[0070] 3) Shuffle the N according to the random position index. R A simulated discrete noise point is assigned to a specified location in the image;

[0071] Step S4.5: Generate analog discrete noise points for Gr / Gb / B channels according to steps S4.1 to S4.4.

[0072] The method described is applicable to the generation of training data for image denoising neural networks.

[0073] Therefore, this solution has the following advantages:

[0074] Discrete noise that exceeds the range of Gaussian distribution is modeled separately, making noise modeling under low light conditions more accurate, reducing the deviation between synthetic data and real noise distribution, and improving the denoising effect of the denoising model in low light environment. Attached Figure Description

[0075] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0076] Figure 1 This is a schematic diagram of the method in this application.

[0077] Figure 2 This is a schematic diagram of the modeling process for discrete noise points in low-light images in this application.

[0078] Figure 3(a) is a schematic diagram of the noise reduction effect before using this solution.

[0079] Figure 3(b) is a comparison with Figure 3(a), showing the denoising effect after using this method. Detailed Implementation

[0080] To better understand the technical content and advantages of the present invention, the present invention will now be described in further detail with reference to the accompanying drawings.

[0081] This invention proposes a method for modeling discrete noise points in low-light images, the overall process of which is as follows: Figure 1 As shown, it includes the following steps:

[0082] S1: Use the target camera to capture multiple black frames and perform black level correction on the black frames; including: capturing RAW format black frames from the target camera in a dark environment, the number of which is multiple, and the number of black frames is determined by the user. In this embodiment, the number of black frames is set to 10.

[0083] Black level correction is performed on black frames, including converting the RAW image to a multi-channel format and subtracting the corresponding black level value from each channel independently.

[0084] S2: Extract discrete noise points from the corrected black frame based on criteria, which include, but are not limited to, the 3σ criterion and the 6σ criterion. This embodiment uses the 3σ criterion; including:

[0085] Calculate the mean μ and standard deviation σ of the pixel values ​​for each channel;

[0086] Pixels with values ​​exceeding the range of μ±3σ are classified as discrete noise.

[0087] S3: Based on the extracted discrete noise points, calculate the probability mass function (PMF) and cumulative distribution function (CDF) for each channel; including:

[0088] The frequency of discrete noise points is statistically analyzed and normalized to obtain the PMF;

[0089] CDF is calculated based on PMF;

[0090] S4: Based on CDF, simulated discrete noise is generated through inverse transform sampling; including:

[0091] The number N of simulated discrete noise points is determined based on actual needs. R ;

[0092] N is generated in a uniform distribution in [0,1]. R A random number U R ;

[0093] The amplitude of simulated discrete noise is generated by mapping random numbers using the inverse function of CDF.

[0094] To assign random spatial locations to simulate discrete noise points;

[0095] Generate simulated discrete noise points for R / Gr / Gb / B channels following the steps described above;

[0096] S5: Output: Generates an image with simulated discrete noise.

[0097] Specifically, the method is as follows Figure 2 As shown, it includes:

[0098] Step S1. Acquire black frames using the target camera and perform black level correction. Black frames are RAW images acquired by the camera in a dark environment, containing all signal-independent noise such as readout noise and stripe noise. Based on black frames, the distribution and amplitude of discrete noise points can be analyzed in detail.

[0099] Step S1.1: Acquire black frames using the target camera. The method for acquiring black frames is to cover the camera lens with black and save the RAW image output by the camera in a completely dark room. To avoid random errors, multiple black frames need to be acquired. The number of black frames can be set according to actual conditions. In this invention, the number of black frames acquired is set to 10.

[0100] Step S1.2 involves performing black level correction on the black frame to eliminate the reference signal offset generated by the image sensor under no-light conditions and restore the true image. Considering that the dark current of different color channels (such as R / Gr / Gb / B) may be inconsistent, each channel is adjusted independently during correction. Assuming the Bayer format of the RAW image is RGGB, the single-channel Bayer format RAW image is first converted into a four-channel image, and the corresponding black level is subtracted from each channel to obtain the black level corrected image.

[0101] X R =X R -BL R (1)

[0102] X Gr =X Gr -BL Gr (2)

[0103] X Gb =X Gb -BL Gb (3)

[0104] X B =X B -BL B (4)

[0105] Step S2. Extract discrete noise based on the black frame. The readout noise and stripe noise in the black frame generally follow a standard Gaussian distribution. The amplitude of discrete noise often exceeds the range represented by the Gaussian distribution; therefore, discrete noise can be considered as outliers in the black frame. Taking the R channel of the black frame as an example, using X... R ={x1,x2,...,x n} represents all pixels in the R channel of the black frame, where n is the total number of pixels in the R channel. Outliers in the black frame are extracted using the 3σ criterion from statistics. The specific steps are as follows:

[0106] Step S2.1, calculate X R mean μ Rand standard deviation σ R ;

[0107]

[0108] Step S2.2, extract X based on the 3σ criterion R outlier X o As discrete noise, the criteria for extracting discrete noise can be determined according to the actual situation, including but not limited to the 3σ criterion and the 6σ criterion;

[0109] X R_o =(X R >μ R +3σ R )∪(X R <μ R -3σ R (7)

[0110] Step S2.3: Extract the discrete noise points of the Gr / Gb / B channels according to steps S2.1 and S2.2 to obtain X. Gr_o X Gb_o and X G_o ;

[0111] X Gr_o =(X Gr >μ Gr +3σ Gr )∪(X Gr <μ Gr -3σ Gr (8)

[0112] X Gb_o =(X Gb >μ Gb +3σ Gb )∪(X Gb <μ Gb -3σ Gb (9)

[0113] X G_o =(X G >μ G +3σ G )∪(X G <μ G -3σ G (10)

[0114] Step S3. Extract discrete noise X from the R / Gr / Gb / B channels R_o X Gr_o X Gb_o and X G_o Calculate the probability mass function and cumulative distribution function for each channel.

[0115] Step S3.1: The distribution of discrete noise points is discrete. The probability mass function can be obtained by calculating the frequency histogram. The calculation steps are as follows:

[0116] 3) Statistics X R_o ={x1,x2,...,x k The number of times each value appears in the expression {m1, m2, ..., m} k};

[0117] 4) Normalize the frequency to obtain the probability mass function P(X=x) of the discrete noise point. k );

[0118]

[0119] 3) Calculate the probability mass function of the Gr / Gb / B channel according to steps 1) and 2).

[0120] Step S3.2: Calculate the corresponding cumulative distribution function based on the probability mass function of each channel. The specific steps are as follows:

[0121] 4) Convert discrete noise point X R_o All possible values ​​are arranged in ascending order: x a <x b <... <x n ;

[0122] 5) For each value x i Calculate F(x) i ) = P R (X≤x i That is, accumulating from x a To x i All probability mass function values;

[0123] 6) After calculating all values, the cumulative distribution function F is obtained. R (x)=P R (X≤x);

[0124] 7) Calculate the cumulative distribution function of the Gr / Gb / B channels according to steps 1), 2), and 3).

[0125] Step S4. Based on the cumulative distribution function, the inverse transform sampling method is used to generate simulated discrete noise points. The specific steps are as follows:

[0126] Step S4.1: Calculate the number of simulated discrete noise points. Taking the R channel as an example, in this embodiment, the total number of pixels in the R channel is M. R =540*960=518400, discrete noise point X R_o The total number of pixels is K R=2436, the number of discrete noise points N to be generated R The calculation method is as follows:

[0127]

[0128] In equation (12), H and W represent the width and height of the simulated noise image, respectively, and α represents a coefficient controlling the number of discrete noise points. The simulated noise image refers to a noisy image obtained by adding simulated noise to a clean image, which is then paired with the clean image to form an image pair for training a denoising neural network. In this embodiment, H = W = 256, α = 2, and the number of discrete noise points generated by the R channel is... Step S4.2, generate N R random numbers Step S4.3, for each random number u i Search for P R (X≤x i )≥u i The smallest x i The value is N, obtained after searching for all random numbers. R A simulated discrete noise point;

[0129] Step S4.4: The location of the discrete noise points in the image is random. After generating the discrete noise points, the number is N. R The simulated discrete noise points are assigned random spatial locations, and the specific steps are as follows:

[0130] 4) Randomly select N from the range of 0 to H*W R 1 integer as the position index;

[0131] 5) N R The discrete noise points of the simulation are randomly shuffled;

[0132] 6) Shuffle the N based on the random position index. R A simulated discrete noise point is assigned to a specified location in the image.

[0133] Step S4.5: Generate simulated discrete noise points for the Gr / Gb / B channels according to steps S4.1 to S4.3. Step S5: Output the simulation results.

[0134] In summary, this invention proposes a method for modeling discrete noise points in low-light images, the main components of which include:

[0135] Step S1. Acquire black frames using the target camera and perform black level correction. Black frames are RAW images acquired by the camera in a dark environment, containing all signal-independent noise such as readout noise and stripe noise. Based on black frames, the distribution and amplitude of discrete noise points can be specifically analyzed.

[0136] Step S2. Extract discrete noise based on black frames. The readout noise and stripe noise in black frames basically follow a standard Gaussian distribution. The amplitude of discrete noise often exceeds the range represented by the Gaussian distribution. Therefore, discrete noise can be regarded as outliers in black frames, and outliers in black frames can be extracted using statistical criteria.

[0137] Step S3. Extract discrete noise X from the R / Gr / Gb / B channels R_o X Gr_o X Gb_o and X G_o Calculate the probability mass function and cumulative distribution function for each channel.

[0138] Step S4. Based on the cumulative distribution function of the R / Gr / Gb / B channels, random numbers uniformly distributed in the 0-1 interval are sampled through inverse transformation to generate simulated discrete noise.

[0139] As shown in Figures 3(a) and 3(b), these are before and after images of using this solution. It can be seen that the noise reduction effect is significant, and the effect after using this solution is obviously better and clearer.

[0140] This technical solution can be applied to Ingenic's series of chips and similar chips that support AI noise reduction.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for modeling discrete noise points in low-light images, characterized in that, The method includes the following steps: S1: Acquire multiple black frames using the target camera, and perform black level correction on the black frames; including: Multiple RAW format black frames were captured from the target camera in a dark environment. Black level correction is performed on black frames, including converting the RAW image to a multi-channel format and subtracting the corresponding black level value from each channel independently. S2: Extract discrete noise points from the corrected black frame based on criteria, including but not limited to the 3σ and 6σ criteria; including: Calculate the mean μ and standard deviation σ of the pixel values ​​for each channel; Pixels with values ​​exceeding the range of μ±3σ are classified as discrete noise. S3: Based on the extracted discrete noise points, calculate the probability mass function (PMF) and cumulative distribution function (CDF) for each channel; including: The frequency of discrete noise points is statistically analyzed and normalized to obtain the PMF; CDF is calculated based on PMF; S4: Based on CDF, simulated discrete noise is generated through inverse transform sampling; including: The number N of simulated discrete noise points is determined based on actual needs. R ; N is generated in a uniform distribution in [0,1]. R A random number U R ; The amplitude of simulated discrete noise is generated by mapping random numbers using the inverse function of CDF. To assign random spatial locations to simulate discrete noise points; Generate simulated discrete noise points for R / Gr / Gb / B channels following the steps described above; S5: Output: Generates an image with simulated discrete noise.

2. The method for modeling discrete noise points in low-light images according to claim 1, characterized in that, Step S1: Acquire multiple black frames using the target camera and perform black level correction on the black frames; the black frames are RAW images acquired by the camera in a dark environment, containing all signal-independent noise such as readout noise and stripe noise. Based on the black frames, the distribution and amplitude of discrete noise points can be specifically analyzed; including: S1.1: Acquire multiple RAW format black frames from the target camera in a dark environment. The black frames are acquired by covering the camera lens and saving the RAW image output by the camera in a completely dark room. To avoid random errors, multiple black frames need to be acquired. The number of black frames can be set according to actual conditions. For example, the number of black frames is set to 10. S1.2: Perform black level correction on black frames, including converting the RAW image to a multi-channel format and subtracting the corresponding black level value from each channel independently; Black frames are corrected for black levels to eliminate the reference signal offset generated by the image sensor under no-light conditions and restore the true image. Considering that the dark current of different color channels, including R / Gr / Gb / B, may be inconsistent, each channel is adjusted independently during correction. Assuming the Bayer format of the RAW image is RGGB, first convert the single-channel Bayer format RAW image into a four-channel image, and subtract the corresponding black level from each channel to obtain the black level corrected image; X R =X R -BL R (1) X Gr =X Gr -BL Gr (2) X Gb =X Gb -BL Gb (3) X B =X B -BL B (4).

3. The method for modeling discrete noise points in a low-light image according to claim 1, characterized in that, In step S2, discrete noise points are extracted based on the black frame. The readout noise and stripe noise in the black frame basically follow a standard Gaussian distribution. The amplitude of discrete noise points often exceeds the characterization range of the Gaussian distribution. Therefore, discrete noise points can be regarded as outliers in the black frame. Assuming the R channel is a black frame, use X R ={x1,x2,...,x n } represents all pixels in the R channel of the black frame, and n is the total number of pixels in the R channel. Assuming that the 3σ criterion in statistics is used to extract outliers in the black frame, the specific steps are as follows: Step S2.1, calculate X R mean μ R and standard deviation σ R ; Step S2.2, assuming X is extracted based on the 3σ criterion R outlier X o As discrete noise, the criteria for extracting discrete noise can be determined according to the actual situation; X R_o =(X R >m R +3s R )∪(X R <m R -3s R ) (7) Step S2.3: Extract the discrete noise points of the Gr / Gb / B channels according to steps S2.1 and S2.2 to obtain X. Gr_o X Gb_o and X G_o ; X Gr_o =(X Gr >m Gr +3s Gr )∪(X Gr <m Gr -3s Gr ) (8) X Gb_o =(X Gb >m Gb +3s Gb )∪(X Gb <m Gb -3s Gb ) (9) X G_o =(X G >m G +3s G )∪(X G <m G -3s G ) (10).

4. The method for modeling discrete noise points in a low-light image according to claim 1, characterized in that, In step S3, the discrete noise point X extracted from the R / Gr / Gb / B channels is... R_o X Gr_o X Gb_o and X G_o Calculate the probability mass function and cumulative distribution function for each channel; Step S3.1: The distribution of discrete noise points is discrete. The probability mass function (PMF) can be obtained by calculating the frequency histogram. The calculation steps are as follows: 1) Statistics X R_o ={x1,x2,...,x k The number of times each value appears in the expression {m1, m2, ..., m} k }; 2) Normalize the frequency to obtain the probability mass function P(X=x) of the discrete noise point. k ); 3) Calculate the probability mass function of the Gr / Gb / B channel according to steps 1) and 2); Step S3.2: Calculate the corresponding cumulative distribution function (CDF) based on the probability mass function of each channel; the specific steps are as follows: 1) Discrete noise point X R_o All possible values ​​are arranged in ascending order: x a <x b <... <x n ; 2) For each value x i Calculate F(x) i ) = P R (X≤x i That is, accumulating from x a To x i All probability mass function values; 3) After calculating all values, the cumulative distribution function F is obtained. R (x)=P R (X≤x); Calculate the cumulative distribution function of the Gr / Gb / B channels according to steps 1), 2), and 3).

5. The method for modeling discrete noise points in a low-light image according to claim 1, characterized in that, Step S4 involves generating simulated discrete noise points using an inverse transform sampling method based on the cumulative distribution function. The specific steps are as follows: Step S4.1: Calculate the number of simulated discrete noise points; assuming it's the R channel, the total number of pixels in the R channel is M. R Discrete noise X R_o The total number of pixels is K R The number of discrete noise points N to be generated R The calculation method is as follows: In equation (12), H and W represent the width and height of the simulated noise image, respectively, and α represents the coefficient that controls the number of discrete noise points. The simulated noise image refers to a noisy image obtained by adding simulated noise to a clean image, and then forming an image pair with the clean image for training a denoising neural network. Step S4.2, generate N R random numbers Step S4.3, for each random number u i Search for P R (X≤x i )≥u i The smallest x i The value is N, obtained after searching for all random numbers. R A simulated discrete noise point; Step S4.4: The location of the discrete noise points in the image is random. After generating the discrete noise points, the number is N. R The simulated discrete noise points are assigned random spatial locations, and the specific steps are as follows: 1) Randomly select N from the range of 0 to H*W R 1 integer as the position index; 2) N R The discrete noise points of the simulation are randomly shuffled; 3) Shuffle the N according to the random position index. R A simulated discrete noise point is assigned to a specified location in the image; Step S4.5: Generate analog discrete noise points for Gr / Gb / B channels according to steps S4.1 to S4.

4.

6. The method for modeling discrete noise points in a low-light image according to claim 1, characterized in that, The method described is applicable to the generation of training data for image denoising neural networks.

Images