A fixed pattern noise (FPN) modeling method for a CMOS image sensor

Abstract

The application provides a fixed pattern noise FPN modeling method of a CMOS image sensor, and the method comprises the following steps: S1, acquiring FPN; S2, FPN statistical analysis; S3, FPN distribution test; S4, establishing a continuous-discrete joint distribution model of FPN; and S5, synthesizing simulated FPN. In the method, only black frames under different ISOs need to be collected for each CMOS image sensor to establish a unified FPN noise model, and the method has strong universality; and the FPN is modeled as a continuous-discrete joint distribution, and the matching degree of the single distribution with the real FPN distribution is higher.

Description

Technical Field

[0001] This invention belongs to the field of image noise reduction technology, and specifically relates to a fixed-mode noise FPN modeling method for CMOS image sensors. Background Technology

[0002] In existing technologies, CMOS image sensors inevitably introduce various noise sources during the imaging process, mainly including photon scattering noise, thermal noise, fixed-pattern noise, readout noise, and quantization noise. Fixed-pattern noise (FPN) is a common type of noise in CMOS image sensors. It exists in a fixed spatial pattern across consecutive image frames, typically manifesting as numerous independent point-like noises and spatially continuous, non-uniform shadows. FPN significantly reduces the signal-to-noise ratio of an image, especially under low-light conditions, where the noise becomes more pronounced and severely degrades image quality. Therefore, analyzing the composition and statistical characteristics of FPN and establishing an accurate noise model to eliminate it is of great significance for improving image quality in low-light environments.

[0003] For photon scattering noise, thermal noise, readout noise, and quantization noise in CMOS image sensors, there are currently relatively mature noise models. For example, Poisson distribution is used to model photon scattering noise, Gaussian distribution is used to model thermal noise, Gaussian or Tukey-lambda distribution is used to model readout noise, and uniform distribution is used to model quantization noise.

[0004] With the development and maturation of deep learning technology, generative techniques such as generative adversarial networks have also been used to build implicit noise models.

[0005] However, current technologies do not establish mathematical models based on the statistical characteristics of FPN. For the numerous noise components in CMOS sensors, existing technologies have established mathematical models for most noise components using a single distribution, achieving high accuracy. Because the components of FPN are complex and do not conform to a single known theoretical distribution, current technologies have not yet addressed the establishment of FPN noise models. Since generative network models rely on a large number of samples to learn the distribution characteristics of noise, and FPN has the characteristic of spatial pattern invariance, it is difficult to obtain sufficiently diverse FPN samples; therefore, the modeling accuracy of generative models is difficult to guarantee.

[0006] In addition, commonly used technical terms include:

[0007] CMOS Image Sensor: A CMOS image sensor (CIS) is an image sensor manufactured using CMOS (Complementary Metal-Oxide-Semiconductor) technology. Due to its advantages such as high integration, low power consumption, and high speed, this type of sensor is widely used in various digital cameras, mobile phones, surveillance equipment, and other fields.

[0008] FPN: Fixed Pattern Noise (FPN) is a common type of noise in image sensors, characterized by a fixed spatial noise distribution in the image, meaning that the location and pattern of the noise remain unchanged each time the same image is captured. This type of noise is particularly noticeable in CMOS image sensors.

[0009] Black frame: An image output by an image sensor in the absence of light.

[0010] ISO: In the field of photography, ISO (International Organization for Standardization) is a standard for measuring the sensitivity of a camera, often referred to as ISO sensitivity or ISO speed. In digital cameras, ISO controls the sensitivity of the camera sensor to light, thus affecting the brightness and noise level of the image. Summary of the Invention

[0011] To address the aforementioned issues, the purpose of this application is to establish an accurate fixed-mode noise model based on the analysis of the statistical characteristics of fixed-mode noise in CMOS image sensors. By accurately simulating the FPN of different CMOS image sensors, simulated noise images with higher similarity to real noise images can be generated on a large scale. This model can be used to train deep learning-based image denoising models, thereby improving the denoising performance of the models in real-world scenarios.

[0012] Specifically, the present invention provides a fixed-mode noise FPN modeling method for CMOS image sensors, the method comprising the following steps:

[0013] S1. Obtaining FPN: Based on the assumption that other noises besides FPN have zero mean characteristics, a fixed pattern noise is obtained by collecting black frames containing only signal-independent noise in a dark environment, and by superimposing and averaging the black frames.

[0014] S2.FPN Statistical Analysis: Based on the FPN acquired by S cameras equipped with the target CMOS image sensor at various ISOs, statistical histograms are plotted. Assuming that the FPN at ISO=15500 is selected for display, the statistical distribution of different cameras equipped with the same type of CMOS image sensor tends to be consistent, indicating that for the same type of CMOS image sensor, a unified mathematical model can be used to represent FPN.

[0015] S3. FPN Distribution Test: The statistical histogram shows that the FPN distribution generally exhibits a bell-shaped distribution with a significant long-tail property. Therefore, the Gaussian and Tukey-lambda distributions are chosen to test the FPN distribution to determine whether it follows a known theoretical distribution. The results of the Gaussian and Tukey-lambda distribution tests are used to determine the distribution. 2 It can be seen that the FPN has a higher degree of matching with the Tukey-lambda distribution than the Gaussian distribution. From the results of the Tukey-lambda distribution, it can be observed that the main part of the FPN has a high degree of matching with the Tukey-lambda distribution, while the tail deviates significantly from the fitted straight line, indicating that it does not basically follow the Tukey-lambda distribution.

[0016] S4. Establish a continuous-discrete joint distribution model for FPN; Based on the FPN distribution test results in step S3, it can be seen that the distributions of the main body and tail of FPN differ significantly and are not limited to following a single distribution. Therefore, appropriate probability distributions are selected for modeling the main body and tail respectively. Since the main body has a high degree of agreement with the Tukey-lambda distribution, the Tukey-lambda distribution is used for modeling; the tail has a low degree of agreement with both the Gaussian distribution and the Tukey-lambda distribution, and the distribution is relatively sparse, so a discrete distribution is suitable for modeling.

[0017] S5. Synthesize and simulate FPN; simulate FPN according to the established continuous-discrete joint distribution; for any ISO, generate a noise image of a specified size that follows the Tukey-lambda distribution according to the calibrated shape parameter λ, scale parameter σ and position parameter μ, and randomly generate discrete noise points according to the discrete probability distribution established under the corresponding ISO and add them to the noise image to obtain the final simulated noise image.

[0018] Step S1 further includes:

[0019] S1.1 Black Frame Acquisition; Prepare S cameras equipped with the target CMOS image sensor. For each camera, select ISO sequentially within the ISO setting range. Acquire N black frames at each ISO setting. Theoretically, the more black frames acquired, the smaller the FPN error.

[0020] S1.2 Calculate FPN; Based on the black frames collected under multiple ISOs, obtain the corresponding FPN for each ISO according to the calculation method of Equation (1);

[0021]

[0022] In formula (1), FPN (ISO) The FPN calculated under a specific ISO. It is the i-th frame among N black frames under a specific ISO.

[0023] In step S1.1, the ISO range of the camera is selected sequentially from 1000 to 15500, so the ISO is set sequentially to 1000, 2000, 4000, 8000, and 15500; assuming N = 500 is set.

[0024] In step S3, the Gaussian and Tukey-lambda distributions are selected to test the distribution of the FPN and determine whether the FPN follows a known theoretical distribution. Since other bell-shaped distributions, including the Cauchy and t distributions, have heavy tails, which do not match the light-tailed characteristics shown by the FPN statistical histogram, the Gaussian and Tukey-lambda distributions are chosen. The testing steps are as follows:

[0025] 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n)};

[0026] 2) Calculate the quantile Q of the corresponding theoretical distribution. i , i≤n;

[0027] 3) Sort the data X (i) and the corresponding theoretical distribution quantile Q i The resulting pairs of data points are plotted as scattered points on the coordinate plane. If the FPN is similar to the selected theoretical distribution, all data points tend to fall on the y=x line; if the two distributions are linearly correlated, the data points tend to fall on a straight line, but not necessarily on the y=x line.

[0028] Step S4 further includes:

[0029] S4.1. Divide the FPN into its main body and tail; the tail of the FPN can be considered as an outlier outside the main distribution. Based on this assumption, the 3σ criterion can be used to divide the FPN into two parts: the main body and the tail. The division method is as follows:

[0030] FPN body = {μ-3σ≤FPN≤ μ+3σ} Formula (2)

[0031] FPN tail = {(FPN≤ μ-3σ)∪FPN≥ μ+3σ} Formula (3)

[0032] S4.2. Modeling the Main Body of the FPN; Since the main body of the FPN closely matches the Tukey-lambda distribution, this distribution is directly used to model the main body of the FPN; For each ISO, the shape parameter λ of the Tukey lambda distribution is estimated using the maximum likelihood estimation (MLE) method. The probability density function of the Tukey lambda distribution is:

[0033]

[0034] Where μ and σ are the location parameter and the proportionality parameter of the distribution, respectively, and the log-likelihood function is:

[0035]

[0036] Differentiate the log-likelihood function with respect to λ, and solve the equation that makes equation (5) equal to 0 to find an estimate of λ;

[0037]

[0038] After obtaining the shape parameter λ of the Tukey-lambda distribution, solve for the scale parameter σ and position parameter μ of the Tukey-lambda distribution according to the following steps:

[0039] 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n)};

[0040] 2) Calculate the quantile Q of the corresponding Tukey-lambda distribution. i , i≤n;

[0041] 3) Establish X using least squares fitting. (i) and Q i Linear relationship:

[0042] X (i) =k·Q i +b Equation (7)

[0043] The slope k and intercept b obtained by least squares fitting can be used as approximations of the scale parameter σ and location parameter μ of the Tukey-lambda distribution.

[0044] S4.3. FPN Tail Modeling: First, draw a statistical histogram of the FPN tail data. It can be observed that the distribution trend of the FPN tail data of different individuals equipped with the same type of CMOS image sensor is consistent, and it has the characteristics of sparse distribution and low probability of occurrence, which is suitable for a unified discrete probability distribution representation.

[0045] In step S4.3, assuming a single ISO, the modeling steps are as follows:

[0046] 1) Calculate the FPN tail data acquired by each camera at each ISO in different value ranges [Z1, Z2, ... Z]. b The probabilities within the range [p1, p2, ... p] are [p1, p2, ... p]. b-1 ], where b is the number of intervals;

[0047] 2) Establish discrete probability distributions for the FPN tail data of different individuals of the same type of CMOS image sensor, as shown in Equation (5):

[0048]

[0049] In the above formula, S represents the number of cameras. For each ISO, a discrete probability model for the tail of the FPN is established according to the above steps.

[0050] In the method described, b = 20 is set.

[0051] The method may further include step S6: verifying the effectiveness of the FPN modeling method;

[0052] Simulated noise was generated using Gaussian distribution and the continuous-discrete joint distribution proposed in this method, respectively, and added to the noise-free image to generate "noise image-noise-free image" paired data for training the denoising model. The same training configuration was used to ensure that the data synthesis method was a single variable.

[0053] After training is complete, the two models are tested on test samples.

[0054] The results show that, compared with the Gaussian distribution, the continuous-discrete joint distribution proposed in this method can train a model with better denoising effect, with fewer pseudo textures and more details after denoising.

[0055] Therefore, the advantage of this application is:

[0056] (1) Each CMOS image sensor only needs to collect black frames under different ISOs to establish a unified FPN noise model, which has strong universality.

[0057] (2) Modeling FPN as a continuous-discrete joint distribution has a higher degree of matching with the real FPN distribution compared to a single distribution. Attached Figure Description

[0058] 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.

[0059] Figure 1 This is a flowchart illustrating the method.

[0060] Figure 2 This is the FPN statistical histogram of the same type of CMOS image sensor.

[0061] Figure 3 This is the Gaussian distribution probability plot of FPN.

[0062] Figure 4 This is the Tukey-lambda distribution probability plot of FPN.

[0063] Figure 5 This is the tail statistics histogram of FPN.

[0064] Figure 6 These are the test results of image denoising models trained using Gaussian distribution and our method, respectively, as noise modeling methods. Detailed Implementation

[0065] 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.

[0066] In this embodiment, a method for FPN modeling of a CMOS image sensor is described, with the overall process as follows: Figure 1 As shown, it includes the following steps:

[0067] Step S1. Obtain FPN. Assuming that other noises besides FPN have zero mean characteristics, fixed pattern noise is obtained by collecting black frames containing only signal-independent noise in a dark environment, and by superimposing and averaging the black frames.

[0068] Step S1.1 Black Frame Acquisition. Prepare S cameras equipped with the target CMOS image sensor. In this embodiment, a camera equipped with an OS03A10 CMOS image sensor is selected, and S=3 is set. For each camera, select ISO values ​​sequentially within the ISO settable range of 1000-15500, setting ISO to 1000, 2000, 4000, 8000, and 15500. Acquire N black frames at each ISO setting. Theoretically, the more black frames acquired, the smaller the FPN error. In this example, N=500 is set.

[0069] Step S1.2 Calculate FPN. Based on the black frames collected under multiple ISOs, obtain the corresponding FPN for each ISO according to the calculation method of Equation (1).

[0070]

[0071] In formula (1), FPN (ISO) The FPN calculated under a specific ISO. It is the i-th frame among N black frames under a specific ISO.

[0072] Step S2. FPN Statistical Analysis. Based on the FPN acquired by S cameras equipped with the target CMOS image sensor at various ISOs, a statistical histogram is plotted. Here, the FPN at ISO=15500 is selected for display. Figure 2 As shown, sensors 1, 2, and 3, i.e., (a), (b), and (c) in the figure, are the same type of CMOS image sensors. For example, sensors 1, 2, and 3 are all of the OS03A10 model. Figure 2 As can be seen, for different cameras equipped with the same type of CMOS image sensor, their statistical distribution tends to be consistent, indicating that for the same type of CMOS image sensor, a unified mathematical model can be used to represent FPN.

[0073] Step S3. FPN distribution test. From Figure 2 The statistical histogram shown indicates that the FPN distribution generally exhibits a bell-shaped distribution with a relatively obvious long-tailed property. The Gaussian and Tukey-lambda distributions are chosen to test the FPN distribution and determine whether it follows a known theoretical distribution. Since other bell-shaped distributions (such as the Cauchy and t-distributions) have heavy tails, which does not match the light-tailed property exhibited by the FPN statistical histogram, the Gaussian and Tukey-lambda distributions are selected. The testing steps are as follows:

[0074] 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n)};

[0075] 2) Calculate the quantile Q of the corresponding theoretical distribution. i , i≤n;

[0076] 3) Sort the data X (i) and the corresponding theoretical distribution quantile Q i The resulting pairs of data points are plotted as scattered points on the coordinate plane. If the FPN is similar to the selected theoretical distribution, all data points tend to fall on the y=x line; if the two distributions are linearly correlated, the data points tend to fall on a straight line, but not necessarily on the y=x line.

[0077] The results of testing whether FPN follows a Gaussian distribution and a Tukey-lambda distribution are as follows: Figure 3 and Figure 4 As shown. The coefficient of determination R from the test results... 2 It can be seen that FPN has a higher degree of matching with the Tukey-lambda distribution compared to the Gaussian distribution. From... Figure 4 It can be observed that the main part of FPN is highly consistent with the Tukey-lambda distribution, while the tail deviates significantly from the fitted straight line, indicating that it does not basically follow the Tukey-lambda distribution.

[0078] Step S4. Establish a continuous-discrete joint distribution model for the FPN; based on the FPN distribution test results in Step S3, it can be seen that the distributions of the main body and tail of the FPN differ significantly and are not limited to following a single distribution. Therefore, appropriate probability distributions are selected for modeling the main body and tail respectively. Since the main body has a high degree of agreement with the Tukey-lambda distribution, the Tukey-lambda distribution is used for modeling; the tail has a low degree of agreement with both the Gaussian and Tukey-lambda distributions, and the distribution is relatively sparse, making a discrete distribution suitable for modeling.

[0079] Step S4.1. Divide the FPN into the main body and tail. The tail of the FPN can be regarded as an outlier outside the main distribution. Based on this assumption, the 3σ criterion can be used to divide the FPN into two parts: the main body and the tail. The division method is as follows:

[0080] FPN body = {μ-3σ≤FPN≤ μ+3σ} Formula (2)

[0081] FPN tail = {(FPN≤ μ-3σ)∪FPN≥ μ+3σ} Formula (3)

[0082] Step S4.2. Modeling the Main Body of the FPN. Since the main body of the FPN closely matches the Tukey-lambda distribution, this distribution is directly used to model the main body of the FPN. For each ISO, the shape parameter λ of the Tukey lambda distribution is estimated using the maximum likelihood estimation (MLE) method. The probability density function of the Tukey lambda distribution is:

[0083]

[0084] Where μ and σ are the location parameter and the proportionality parameter of the distribution, respectively, and the log-likelihood function is:

[0085]

[0086] Differentiate the log-likelihood function with respect to λ, and solve the equation with equation (5) set to 0 to find an estimate of λ.

[0087]

[0088] After obtaining the shape parameter λ of the Tukey-lambda distribution, solve for the scale parameter σ and position parameter μ of the Tukey-lambda distribution according to the following steps:

[0089] 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n)};

[0090] 2) Calculate the quantile Q of the corresponding Tukey-lambda distribution. i , i≤n;

[0091] 3) Establish X using least squares fitting. (i) and Q i Linear relationship:

[0092] X (i) =k·Q i +b Equation (7)

[0093] The slope k and intercept b obtained by least-squares fitting can be used as approximations of the scale parameter σ and location parameter μ of the Tukey-lambda distribution.

[0094] Step S4.3. FPN Tail Modeling. First, draw a statistical histogram of the FPN tail data, such as... Figure 5 As shown, it can be observed that the distribution trend of the FPN tail data of different individuals equipped with the same type of CMOS image sensor is basically consistent, and it has the characteristics of sparse distribution and low occurrence probability, making it suitable for a unified discrete probability distribution representation. Taking a single ISO as an example, the modeling steps are as follows:

[0095] 1) Calculate the FPN tail data acquired by each camera at each ISO in different value ranges [Z1, Z2, ... Z]. b The probabilities within the range [p1, p2, ... p] are [p1, p2, ... p]. b-1 ], where b is the number of intervals, and in this embodiment, b = 20;

[0096] 2) Establish discrete probability distributions for the FPN tail data of different individuals of the same type of CMOS image sensor, as shown in Equation (8):

[0097]

[0098] In the above formula, S represents the number of cameras. For each ISO, a discrete probability model for the tail of the FPN is established according to the above steps.

[0099] Step S5. Synthesize and simulate the FPN; simulate the FPN based on the established continuous-discrete joint distribution. For any ISO, generate a noise image of a specified size that follows a Tukey-lambda distribution based on the calibrated shape parameter λ, scale parameter σ, and position parameter μ. Randomly generate discrete noise points according to the discrete probability distribution established under the corresponding ISO and add them to the noise image to obtain the final simulated noise image.

[0100] Step S6. Verify the effectiveness of the FPN modeling method; simulated noise is generated using Gaussian distribution and the continuous-discrete joint distribution proposed in this application, respectively, and added to the noise-free image to generate "noisy image-noisy image" paired data for training the denoising model. The training configuration is the same to ensure that the data synthesis method is a single variable. After training is completed, the two models are tested on the test samples, and the test results are as follows. Figure 6 As shown in the figure. The results show that, compared with the Gaussian distribution, using the continuous-discrete joint distribution proposed in this application as the noise simulation model can train a model with better denoising effect, with fewer pseudo textures and more details after denoising.

[0101] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0102] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0103] 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 fixed-mode noise FPN modeling method for CMOS image sensors, characterized in that, The method includes the following steps: S1. Obtaining FPN: Based on the assumption that other noises besides FPN have zero mean characteristics, a fixed pattern noise is obtained by collecting black frames containing only signal-independent noise in a dark environment, and by superimposing and averaging the black frames. S2.FPN Statistical Analysis: Based on the FPN acquired by S cameras equipped with the target CMOS image sensor at various ISOs, statistical histograms are plotted. Assuming that the FPN at ISO=15500 is selected for display, the statistical distribution of different cameras equipped with the same type of CMOS image sensor tends to be consistent, indicating that for the same type of CMOS image sensor, a unified mathematical model can be used to represent FPN. S3. FPN Distribution Test: The statistical histogram shows that the FPN distribution generally exhibits a bell-shaped distribution with a significant long-tail property. Therefore, the Gaussian and Tukey-lambda distributions are chosen to test the FPN distribution to determine whether it follows a known theoretical distribution. The results of the Gaussian and Tukey-lambda distribution tests are used to determine the distribution. 2 It can be seen that the FPN has a higher degree of matching with the Tukey-lambda distribution than the Gaussian distribution. From the results of the Tukey-lambda distribution, it can be observed that the main part of the FPN has a high degree of matching with the Tukey-lambda distribution, while the tail deviates significantly from the fitted straight line, indicating that it does not basically follow the Tukey-lambda distribution. S4. Establish a continuous-discrete joint distribution model for FPN; Based on the FPN distribution test results in step S3, it can be seen that the distributions of the main body and tail of FPN differ significantly and are not limited to following a single distribution. Therefore, appropriate probability distributions are selected for modeling the main body and tail respectively. Since the main body has a high degree of agreement with the Tukey-lambda distribution, the Tukey-lambda distribution is used for modeling; the tail has a low degree of agreement with both the Gaussian distribution and the Tukey-lambda distribution, and the distribution is relatively sparse, so a discrete distribution is suitable for modeling. S5. Synthesize and simulate FPN; simulate FPN according to the established continuous-discrete joint distribution; for any ISO, generate a noise image of a specified size that follows the Tukey-lambda distribution according to the calibrated shape parameter λ, scale parameter σ and position parameter μ, and randomly generate discrete noise points according to the discrete probability distribution established under the corresponding ISO and add them to the noise image to obtain the final simulated noise image.

2. The fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 1, characterized in that, Step S1 further includes: S1.1 Black Frame Acquisition; Prepare S cameras equipped with the target CMOS image sensor. For each camera, select ISO sequentially within the ISO setting range. Acquire N black frames at each ISO setting. Theoretically, the more black frames acquired, the smaller the FPN error. S1.2 Calculate FPN; Based on the black frames collected under multiple ISOs, obtain the corresponding FPN for each ISO according to the calculation method of Equation (1); In formula (1), FPN (ISO) The FPN calculated under a specific ISO. It is the i-th frame among N black frames under a specific ISO.

3. The fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 2, characterized in that, In step S1.1, the ISO range of the camera is selected sequentially from 1000 to 15500, so the ISO is set sequentially to 1000, 2000, 4000, 8000, and 15500; assuming N = 500 is set.

4. The fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 1, characterized in that, In step S3, the Gaussian and Tukey-lambda distributions are selected to test the distribution of the FPN and determine whether the FPN follows a known theoretical distribution. Since other bell-shaped distributions, including the Cauchy and t distributions, have heavy tails, which do not match the light-tailed characteristics shown by the FPN statistical histogram, the Gaussian and Tukey-lambda distributions are chosen. The testing steps are as follows: 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n) }; 2) Calculate the quantile Q of the corresponding theoretical distribution. i , i≤n; 3) Sort the data X (i) and the corresponding theoretical distribution quantile Q i The resulting pairs of data points are plotted as scattered points on the coordinate plane. If the FPN is similar to the selected theoretical distribution, all data points tend to fall on the y=x line; if the two distributions are linearly correlated, the data points tend to fall on a straight line, but not necessarily on the y=x line.

5. A fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 1, characterized in that, Step S4 further includes: S4.

1. Divide the FPN into the main body and tail; the tail of the FPN can be regarded as an outlier outside the main distribution. Based on this assumption, the 3σ criterion can be used to divide the FPN into two parts: the main body and the tail. The division method is as follows: FPN body = {μ-3σ≤FPN≤ μ+3σ} Equation (2) FPN tail = {(FPN≤ μ-3σ)∪FPN≥ μ+3σ} Equation (3) S4.

2. Modeling the Main Body of the FPN; Since the main body of the FPN closely matches the Tukey-lambda distribution, this distribution is directly used to model the main body of the FPN; For each ISO, the shape parameter λ of the Tukey-lambda distribution is estimated using the maximum likelihood estimation (MLE) method. The probability density function of the Tukey-lambda distribution is: Where μ and σ are the location parameter and the proportion parameter of the distribution, respectively, and the log-likelihood function is: Differentiate the log-likelihood function with respect to λ, and solve the equation that makes equation (5) equal to 0 to find an estimate of λ; After obtaining the shape parameter λ of the Tukey-lambda distribution, solve for the scale parameter σ and position parameter μ of the Tukey-lambda distribution according to the following steps: 1) Sort the data in the FPN. Assume that all pixel values ​​in the FPN image are {X1, X2, ... X}. n After sorting the pixel values, we get {X} (1) <X (2) <... <X (n) }; 2) Calculate the quantile Q of the corresponding Tukey-lambda distribution. i , i≤n; 3) Establish X using least squares fitting. (i) and Q i Linear relationship: X (i) =k·Q i +b Equation (7) The slope k and intercept b obtained by least squares fitting can be used as approximations of the scale parameter σ and location parameter μ of the Tukey-lambda distribution. S4.

3. FPN Tail Modeling: First, draw a statistical histogram of the FPN tail data. It can be observed that the distribution trend of the FPN tail data of different individuals equipped with the same type of CMOS image sensor is consistent, and it has the characteristics of sparse distribution and low probability of occurrence, which is suitable for a unified discrete probability distribution representation.

6. The fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 1, characterized in that, In step S4.3, assuming a single ISO, the modeling steps are as follows: 1) Calculate the FPN tail data acquired by each camera at each ISO in different value ranges [Z1, Z2, ... Z]. b The probabilities within the range [p1, p2, ... p] are [p1, p2, ... p]. b-1 ], where b is the number of intervals; 2) Establish discrete probability distributions for the FPN tail data of different individuals of the same type of CMOS image sensor, as shown in equation (5): In the above formula, S represents the number of cameras. For each ISO, a discrete probability model for the tail of the FPN is established according to the above steps.

7. A fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 6, characterized in that, In the method described, b = 20 is set.

8. A fixed-mode noise FPN modeling method for a CMOS image sensor according to claim 6, characterized in that, The method may further include step S6: verifying the effectiveness of the FPN modeling method; Simulated noise was generated using Gaussian distribution and the continuous-discrete joint distribution proposed in this method, respectively, and added to the noise-free image to generate "noise image-noise-free image" paired data for training the denoising model. The same training configuration was used to ensure that the data synthesis method was a single variable. After training is complete, the two models are tested on test samples. The results show that, compared with the Gaussian distribution, the continuous-discrete joint distribution proposed in this method can train a model with better denoising effect, with fewer pseudo textures and more details after denoising.

Images