Noise information generation method and apparatus

Abstract

The embodiment of the application discloses a noise information generation method and device, the method comprises the following steps: obtaining a noise-free image; determining Poisson noise of the noise-free image through a Poisson noise generation model, and determining quasi-elliptical Gaussian noise of the noise-free image through a quasi-elliptical Gaussian noise generation model, the Poisson noise generation model is used for generating noise related to an image signal, and the quasi-elliptical Gaussian noise generation model is used for generating noise irrelevant to the image signal; superimposing the Poisson noise and the quasi-elliptical Gaussian noise into the noise-free image to generate an image containing noise information. The noise generated by the scheme is more real, can provide a large number of noise-containing images conforming to actual conditions for a deep learning noise reduction task, and further improves the robustness of a noise reduction model for identifying different morphological noises.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of noise processing, and particularly relate to a noise information generation method and device. BACKGROUND

[0002] With the popularity of mobile devices and the development of multimedia technology, browsing and watching various multimedia information through mobile devices has become an important function of mobile devices at present, such as watching live video or making video calls through mobile devices. Since noise signals can greatly affect the display effect of the picture, removing noise existing in the picture has always been an important issue in the research and development of video when the mobile device displays the picture.

[0003] In related technologies, when performing noise reduction processing, Gaussian white noise is generally used as a model of synthetic noise, so that the noise size and intensity can be conveniently controlled, and the synthetic Gaussian white noise is uniformly superimposed on a clean image without high-definition noise to obtain a synthetic noise image as a noise reduction input to verify the noise reduction effect. However, there is a difference between Gaussian white noise and actual generated noise, and in the above related noise reduction processing process, due to inaccurate noise modeling, distortion or blurred noise reduction side effects often appear on the image after noise reduction. Therefore, how to generate noise information that is more realistic and conforms to actual objective conditions, so that subsequent model training can obtain a large amount of noise-containing data close to the real distribution as training data to be applied to noise reduction processing is a problem that needs to be solved. SUMMARY

[0004] Embodiments of the present application provide a noise information generation method and device, which solve the problem of low accuracy of synthetic noise in related technologies, can provide a large number of noise-containing images that conform to actual conditions for deep learning noise reduction tasks, and further improve the robustness of the noise reduction model for identifying different forms of noise.

[0005] In a first aspect, embodiments of the present application provide a noise information generation method, which comprises:

[0006] obtaining a noise-free image;

[0007] determining Poisson noise of the noise-free image through a Poisson noise generation model and determining class-elliptical Gaussian noise of the noise-free image through a class-elliptical Gaussian noise generation model, the Poisson noise generation model being used to generate noise related to an image signal, and the class-elliptical Gaussian noise generation model being used to generate noise irrelevant to the image signal;

[0008] superimposing the Poisson noise and the class-elliptical Gaussian noise into the noise-free image to generate an image containing noise information.

[0009] In a second aspect, an embodiment of the present application further provides a noise information generation apparatus, comprising:

[0010] an image acquisition module configured to acquire a noise-free image;

[0011] a noise generation module configured to determine Poisson noise of the noise-free image through a Poisson noise generation model and to determine elliptical-like Gaussian noise of the noise-free image through an elliptical-like Gaussian noise generation model, the Poisson noise generation model being used to generate noise related to an image signal, and the elliptical-like Gaussian noise generation model being used to generate noise irrelevant to the image signal;

[0012] a noise addition module configured to superimpose the Poisson noise and the elliptical-like Gaussian noise into the noise-free image to generate an image containing noise information.

[0013] In a third aspect, an embodiment of the present application further provides a noise information generation device, comprising:

[0014] one or more processors;

[0015] a storage device configured to store one or more programs,

[0016] when the one or more programs are executed by the one or more processors, the one or more processors implement the noise information generation method according to the embodiments of the present application.

[0017] In a fourth aspect, an embodiment of the present application further provides a storage medium storing computer executable instructions, which, when executed by a computer processor, are used to perform the noise information generation method according to the embodiments of the present application.

[0018] In a fifth aspect, an embodiment of the present application further provides a computer program product, which comprises a computer program stored in a computer readable storage medium, and at least one processor of a device reads and executes the computer program from the computer readable storage medium, so that the device performs the noise information generation method according to the embodiments of the present application.

[0019] In the embodiment of the present application, after obtaining the noise-free image, the Poisson noise of the noise-free image is determined through a Poisson noise generation model, and the quasi-elliptical Gaussian noise of the noise-free image is determined through a quasi-elliptical Gaussian noise generation model. The Poisson noise generation model is used to generate noise related to the image signal, and the quasi-elliptical Gaussian noise generation model is used to generate noise unrelated to the image signal. The Poisson noise and the quasi-elliptical Gaussian noise are superimposed on the noise-free image to generate an image containing noise information, so that the generated noise is more realistic, a large number of noise-containing images in line with actual conditions can be provided for the deep learning noise reduction task, and the robustness of the noise reduction model to identify different morphological noises is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of a noise information generation method provided by the embodiment of the present application;

[0021] Figure 2 An image containing noise information obtained by superimposing noise provided by the embodiment of the present application;

[0022] Figure 3 A flowchart of a method for generating Poisson noise corresponding to a noise-free image provided by the embodiment of the present application;

[0023] Figure 4 A flowchart of a method for generating quasi-elliptical Gaussian noise corresponding to a noise-free image provided by the embodiment of the present application;

[0024] Figure 5 A flowchart of a method for constructing a noise model based on a sample image provided by the embodiment of the present application;

[0025] Figure 6 A schematic diagram of a sample image containing a standard color block provided by the embodiment of the present application;

[0026] Figure 7 A structural block diagram of a noise information generation device provided by the embodiment of the present application;

[0027] Figure 8 A structural schematic diagram of a noise information generation device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0028] The embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the embodiments of the present application, but not to limit the embodiments of the present application. In addition, it should be noted that, in order to facilitate description, only parts related to the embodiments of the present application are shown in the drawings, but not all structures.

[0029] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of a kind and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / " generally indicates that the front and rear associated objects are in an "or" relationship.

[0030] Figure 1 A flowchart of a noise information generation method provided by an embodiment of the present application can be used to add noise to an input image to generate an image containing noise information. The method can be executed by a computing device such as a server, a smart terminal, a notebook, a tablet computer, etc., and specifically includes the following steps:

[0031] In step S101, a noise-free image is obtained.

[0032] In an embodiment, noise is added to the image to generate an image containing noise information. The image with added noise information and the original image without added noise information can be used as a sample pair for training a noise reduction processing model, for use in a noise reduction model or other noise reduction processing algorithm.

[0033] First, a noise-free image is obtained. Optionally, the noise-free image can be a clean noise-free video frame image. The noise-free image can be an image that meets certain shooting conditions or certain noise-free quality requirements, to serve as a noise-free image. Optionally, it can also be a clean noise-free image obtained by operating a group of continuous images taken of the same shooting scene. For example, for the same device model at the same set sensitivity, 100 continuous images can be taken, and the clean noise-free image can be generated by calculating the average pixel value of each pixel point of the multiple continuous images and using the average pixel value as the pixel value of the pixel point to generate a noise-free image. For example, for 100 continuous images of a fixed n*m size, for each pixel point, the pixel values of the same pixel point position in the 100 images are added and divided by 100 to obtain the average pixel value of the pixel point position. The average pixel value of each pixel point position is calculated in the same way, and the average pixel value of each pixel point position is used as the pixel value of the position point to generate a noise-free image.

[0034] Step S102, determining the Poisson noise of the noise-free image by a Poisson noise generation model, and determining the elliptical-like Gaussian noise of the noise-free image by an elliptical-like Gaussian noise generation model.

[0035] In one embodiment, after obtaining the noise-free image, the Poisson noise of the noise-free image is determined by a configured Poisson noise generation model, and the elliptical-like Gaussian noise of the noise-free image is determined by a configured elliptical-like Gaussian noise generation model. Optionally, for the obtained noise-free image, the corresponding Poisson noise and elliptical-like Gaussian noise of each pixel point can be calculated, and the calculated Poisson noise and elliptical-like Gaussian noise can be specific pixel values, such as a pixel value of the Poisson noise of a pixel point and a pixel value of the elliptical-like Gaussian noise of the pixel point. For example, for a pixel point in the noise-free image, the pixel value of the pixel point in the noise-free image is x, the pixel value of the Poisson noise of the pixel point calculated by the configured Poisson noise generation model is x1, and the pixel value of the elliptical-like Gaussian noise of the pixel point calculated by the configured elliptical-like Gaussian noise generation model is x2. The Poisson noise generation model is used to generate noise related to the image signal, and the elliptical-like Gaussian noise generation model is used to generate noise unrelated to the image signal, that is, the corresponding noise related to the image signal and the noise unrelated to the image signal in the noise-free image are generated by the Poisson noise generation model and the elliptical-like Gaussian noise generation model respectively.

[0036] Step S103, superimposing the Poisson noise and the elliptical-like Gaussian noise into the noise-free image to generate an image containing noise information.

[0037] In one embodiment, after the Poisson noise and the elliptical-like Gaussian noise corresponding to the noise-free image are determined respectively, the Poisson noise and the elliptical-like Gaussian noise are superimposed into the noise-free image to generate an image containing noise information. Specifically, the original pixel value of each pixel point in the noise-free image can be superimposed with the pixel value after the Poisson noise and the elliptical-like Gaussian noise. For example, the schematic diagram after superimposing the noise is shown in Figure 2 Figure 2 The embodiment of the present application provides an image containing noise information obtained by superimposing the noise. For the actual corresponding heat map, the warmer the color, the higher the noise signal strength.

[0038] Specifically, the noise-free image is denoted as I clean , and the image after superimposing the Poisson noise and the elliptical-like Gaussian noise is denoted as I noise , wherein the Poisson noise is denoted as p', which is subject to Poisson distribution, the Gaussian noise is denoted as go', which is subject to Gaussian distribution, and the calculation process of the generated image containing noise information can be: ​

[0039] I noise =I clean +p′+go′

[0040] In another embodiment, corresponding coefficient factors are set for Poisson noise and elliptical Gaussian noise, respectively. For example, the coefficient factor of Poisson noise is denoted as a, and the coefficient factor of elliptical Gaussian noise is denoted as β. At this time, the image containing noise information can be calculated as follows:

[0041] I noise =I clean +α*p′+β*go′

[0042] Wherein, the coefficient factor represents the expansion or reduction multiple of the noise. For example, a is 2, and β is 3, which respectively represent that the Poisson noise and the elliptical Gaussian noise are expanded by 2 times and 3 times. Through the setting of the coefficient factor, the intensity of the superimposed noise can be controlled, so as to obtain the noise-containing image that meets the requirements of various scenes.

[0043] From the above scheme, when generating an image containing noise information, the Poisson noise of the noise-free image is determined by the Poisson noise generation model, and the elliptical Gaussian noise of the noise-free image is determined by the elliptical Gaussian noise generation model. Then, the two noises are superimposed on the noise-free image, instead of directly using the simple linear superposition of Gaussian noise on the noise-free image as in the related art, so that the generated noise is more realistic, and a large number of noise-containing images that meet the actual situation can be provided for the deep learning noise reduction task, further improving the robustness of the noise reduction model for identifying different morphological noises.

[0044] Figure 3 The flowchart of a method for generating Poisson noise corresponding to a noise-free image provided by the embodiment of the present application is shown in FIG. 1, which specifically includes the following steps. Figure 3

[0045] Step S201, determining the Poisson distribution expectation value of each pixel point according to the pixel value of each pixel point in the noise-free image.

[0046] In one embodiment, when calculating the Poisson noise of the noise-free image, for each pixel point, the Poisson distribution expectation value corresponding to the Poisson distribution is calculated according to the pixel value. The Poisson distribution expectation value and the pixel value of the pixel point satisfy a Poisson mapping function relationship, and the Poisson mapping function relationship is obtained based on polynomial fitting.

[0047] Step S202, calculating the Poisson noise value corresponding to the pixel value of each pixel point through the Poisson distribution formula and the Poisson distribution expectation value of each pixel point. ​

[0048] In an embodiment, after the Poisson distribution expectation value corresponding to each pixel point is determined, the Poisson distribution expectation value and the pixel value (the pixel value as a random variable) of the pixel point are substituted into the Poisson distribution formula to obtain the Poisson noise corresponding to the pixel point, wherein the Poisson distribution expectation value determined based on the pixel point is the expectation satisfied by the Poisson distribution, and the Poisson distribution expectation value is calculated based on the pixel value of the pixel point. For example, the acquired noise-free image is denoted as I input , then the Poisson distribution expectation value λ = f(I input ), that is, the Poisson distribution expectation value λ is a mapping function related to the input signal (the pixel value of the pixel point in the noise-free image), and at this time, the Poisson noise p' ~ Poisson(λ = f(I input ). Optionally, the Poisson mapping function relationship can be a quintic equation or a quartic equation, and the coefficients of the equation are obtained when the model is pre-established. At this time, the result obtained by substituting the pixel value into the equation is the corresponding Poisson distribution expectation value.

[0049] In step S203, the Poisson noise of the noise-free image is generated based on the Poisson noise value of each pixel point.

[0050] In an embodiment, after the Poisson noise value of each pixel point in the noise-free image is calculated, the Poisson noise of the entire image is obtained, that is, the Poisson noise of the entire image is composed of the Poisson noise value of each pixel point.

[0051] As described above, for the noise-free image, when determining the Poisson noise, the pixel value of the pixel point is used to calculate the Poisson distribution expectation value, and after obtaining the Poisson distribution satisfied by each pixel point, the pixel value is substituted to calculate the Poisson noise value. In the process of scheme design, it is found that the noise on different pixel values is not the same, which means that the noise signal depends on the image signal to some extent, and is a signal-dependent noise. In the live broadcast service scenario, due to the differences in actual equipment and lighting (exposure) conditions, the inherent noise will be unevenly distributed on the image. According to the time point change or accumulation, the inherent image signal will further change, and the sampling change of such fluctuation conforms to the Poisson distribution. Therefore, the mapping relationship between the pixel value and the Poisson distribution expectation value is established to calculate the Poisson distribution expectation value through the pixel value, obtain the corresponding Poisson distribution, and then calculate the Poisson noise.

[0052] Figure 4 A flowchart of a method for generating a class-elliptical Gaussian noise corresponding to a noise-free image provided by an embodiment of the present application is shown in Figure 4 , and specifically includes:

[0053] Step S301, determine the pixel distance value of each pixel point in the noise-free image to the image center, and determine the standard deviation of the Gaussian distribution corresponding to each pixel point according to the pixel distance value.

[0054] In one embodiment, when calculating the quasi-elliptical noise of the noise-free image, for each pixel point, the pixel distance value of the pixel point to the image center is determined, and the standard deviation of the Gaussian distribution corresponding to each pixel point is determined according to the pixel distance value. The standard deviation of the Gaussian distribution and the pixel distance value of the pixel point satisfy a Gaussian mapping function relationship, and the Gaussian mapping function relationship is obtained based on polynomial fitting.

[0055] Optionally, the Gaussian mapping function relationship can be a quintic equation or a quartic equation, and the coefficients of the equation are obtained when the model is pre-built. At this time, the result obtained by substituting the pixel distance value into the equation is the standard deviation of the corresponding Gaussian distribution.

[0056] In one embodiment, the center point of the noise-free image is taken as the image center, and a size of H*W is taken as an example. The geometric center (H / 2, W / 2) is equivalent to the image center, which is denoted as (origin h ,origin w ). The straight line distance of each pixel point (a i ,b i ) in the noise-free image to the image center point is calculated one by one.

[0057] Step S302, calculate the Gaussian noise value of each pixel point through the Gaussian distribution formula and the standard deviation of the Gaussian distribution corresponding to each pixel point.

[0058] In one embodiment, after determining the standard deviation of the Gaussian distribution corresponding to each pixel point, the standard deviation of the Gaussian distribution and the pixel distance value (the pixel distance value as a random variable) are substituted into the Gaussian distribution formula, and the Gaussian distribution is a standard Gaussian distribution with an expectation of 0, to obtain the quasi-elliptical noise corresponding to the pixel point, wherein the standard deviation of the Gaussian distribution determined based on the pixel distance value is the standard deviation that the Gaussian distribution obeys, and the standard deviation of the Gaussian distribution is calculated based on the pixel distance value. For example, the standard deviation is denoted as σ oval , and the Gaussian noise go' ~ Gaussian (0, σ oval ).

[0059] Step S303, generate the quasi-elliptical Gaussian noise of the noise-free image based on the Gaussian noise value of each pixel point.

[0060] In one embodiment, after the class-elliptical Gaussian noise value of each pixel in the noise-free image is calculated, the class-elliptical Gaussian noise of the entire image is obtained, that is, the class-elliptical Gaussian noise of the entire image is composed of the class-elliptical Gaussian noise values of each pixel.

[0061] As can be seen from the above, for a noise-free image, when determining the class-elliptical Gaussian noise, the standard deviation of the Gaussian distribution is calculated using the distance value of the pixel point to determine the Gaussian distribution to which the pixel point conforms, and then the pixel distance value is substituted into the Gaussian distribution formula as the independent variable to obtain the corresponding Gaussian noise. In the process of scheme design, considering that after sharpening on a mobile terminal device, the signal-dependent noise attached to the image signal itself will be further amplified in the high frequency field, thus becoming larger noise after the server collects data. At the same time, because of the physical properties of the shooting lens, the number of received photons received by the receiving components is also inconsistent, forming a vignetting effect that is a visible noise signal gradually increasing from the center of the lens to the corners of the picture, presenting a noise distribution effect of high in the middle and low around. And as a kind of persistent random noise, it will be widely distributed under various models and lighting conditions, so the class-elliptical Gaussian noise model based on the pixel distance value is used to generate noise.

[0062] Figure 5 A flowchart of a method for constructing a noise model based on a sample image provided by an embodiment of the present application is shown in Figure 5 , and specifically includes:

[0063] Step S401, obtaining a first sample picture set, the first sample picture set being composed of multiple images continuously shot under the same shooting scene, wherein each image contains one or more noise blocks.

[0064] In one embodiment, the Poisson noise generation model is constructed by using the first sample picture set composed of multiple images continuously shot under the same shooting scene, wherein each image contains one or more noise blocks. For example, the contained noise block can be a standard color card data, for example, as shown in Figure 6 , Figure 6 A schematic diagram of a sample image containing a standard color block provided by an embodiment of the present application. Different square blocks in the middle frame selected area are noise blocks.

[0065] Step S402, obtaining a second sample picture set, the second sample picture set being pictures shot by shielding the camera.

[0066] In one embodiment, when constructing the elliptic Gaussian noise generation model, images captured by an occluded camera are used for processing. If the same camera angle and scene as the first sample image are used, and the camera is covered before shooting, the inherent noise of the system is fully extracted without mixing with image signals and signal-dependent noise. For example, 200 to 300 images may be selected.

[0067] Step S403: Calculate and generate a Poisson noise generation model using the first sample image set, and calculate and generate an elliptic Gaussian noise generation model using the second sample image set.

[0068] In one embodiment, when calculating the Poisson noise generation model using the first sample image set, for each sample image... Calculate the pixel mean and standard deviation of the pixels contained in each noise block, and obtain the Poisson mapping function from the pixel mean to the pixel standard deviation through polynomial fitting. The pixel mean is denoted as Pixel for example. mean It is obtained by summing the pixel values ​​of each pixel in the noise block and then dividing by the number of pixels contained therein, where the noise fluctuation, i.e., the pixel standard deviation, is denoted as std. x Then the mapping relationship between the two can be expressed as std x =f x (Pixel mean ), where λ x =std x That is, the std x The expected value of the Poisson distribution is used to determine the Poisson distribution corresponding to each pixel. The polynomial fitting process involves determining the equation coefficients. For example, a quartic equation is used as the fitting polynomial to simulate the relationship between the actual pixel mean and its corresponding standard deviation. This polynomial is, for example, y = ax². 4 +bx 3 +cx 2 +dx, since the pixel values ​​of the independent variable and the standard deviation of the corresponding dependent variable are both discrete values ​​obtained from sampling statistics, the mapping relationship between them (std) is obtained through the above polynomial fitting method. x =f x (Pixel mean This gives the specific values ​​of a, b, c, and d. It should be noted that the above example uses a quartic equation in one variable; however, it can also be other multivariate equations, which are not limited here.

[0069] In the process of generating the quasi-elliptical Gaussian noise generation model through the second sample picture set: the image center of each picture in the second sample picture set and the noise intensity value of each pixel point are determined, the pixel distance value of each pixel point to the image center is calculated, and the Gaussian mapping function relationship between the pixel distance value and the noise intensity value is calculated through polynomial fitting.

[0070] wherein, for each pixel point (a x ,b x ), it corresponds to a Gaussian distribution of g x (0, σ x ), that is, the quasi-elliptical Gaussian noise of the corresponding pixel point wherein, σ x =std dark , std dark is the noise signal change corresponding to the distance value of different pixel points, which is taken as the standard deviation of the Gaussian distribution.

[0071] The Gaussian mapping function relationship between the pixel distance value and the noise intensity value is calculated through polynomial fitting, wherein the process of polynomial fitting is the process of determining the equation coefficients. For example, a monomial quartic equation is used as a fitting polynomial to simulate the relationship between the actual pixel distance value and the noise intensity, and the polynomial is y = ex 4 +fx 3 +gx 2 +hx. Since the pixel distance value of the independent variable and the noise intensity of the corresponding dependent variable are discrete values obtained by sampling statistics, the mapping relationship between the two is obtained through the above polynomial fitting, that is, the specific values of e, f, g and h are obtained. It should be noted that the above is an example of a monomial quartic equation, which can also be other multivariate equations, which are not limited here.

[0072] As can be seen from the above, the Poisson noise generation model and the quasi-elliptical Gaussian noise generation model are generated by analyzing and processing the sample pictures, which are used for subsequent generation of corresponding Poisson noise and quasi-elliptical noise for noise-free images, so that the generated noise is more realistic, can provide a large number of actual situation containing noise images for deep learning denoising task, and further improves the robustness of the denoising model for identifying different morphological noises.

[0073] On the basis of the technical solution, the first sample picture set and the second sample picture are respectively based on different equipment models and different sensitivities to be captured, and different equipment models and different sensitivities correspond to different Poisson noise generation models and different elliptical Gaussian noise generation models. In the process of generating noise information, for the obtained noise-free image, the shooting equipment corresponding to the noise-free image and the sensitivity parameter when the image is shot are determined, and the corresponding configured Poisson noise generation model and the elliptical Gaussian noise generation model are used to generate Poisson noise and elliptical Gaussian noise respectively and superimposed, so as to further ensure the accuracy of the generated noise information and avoid the problem of inaccurate generated noise caused by shooting with different equipment and different sensitivities.

[0074] Figure 7 A structural block diagram of a noise information generation device provided by an embodiment of the present application is provided, the device is used for executing the noise information generation method provided by the above-mentioned embodiment, and has function modules and beneficial effects corresponding to the execution method. As shown in the figure, the device specifically includes: an image acquisition module 101, a noise generation module 102, and a noise addition module 103, wherein, Figure 7

[0075] The image acquisition module 101 is configured to acquire a noise-free image.

[0076] The noise generation module 102 is configured to determine Poisson noise of the noise-free image through a Poisson noise generation model, and determine elliptical Gaussian noise of the noise-free image through an elliptical Gaussian noise generation model, the Poisson noise generation model is used to generate noise related to an image signal, and the elliptical Gaussian noise generation model is used to generate noise irrelevant to the image signal.

[0077] The noise addition module 103 is configured to superimpose the Poisson noise and the elliptical Gaussian noise into the noise-free image to generate an image containing noise information.

[0078] As known from the above-mentioned solution, after the noise-free image is acquired, the Poisson noise of the noise-free image is determined through the Poisson noise generation model, and the elliptical Gaussian noise of the noise-free image is determined through the elliptical Gaussian noise generation model, wherein the Poisson noise generation model is used to generate noise related to an image signal, and the elliptical Gaussian noise generation model is used to generate noise irrelevant to the image signal, the Poisson noise and the elliptical Gaussian noise are superimposed into the noise-free image to generate an image containing noise information, so that the generated noise is more realistic, a large number of noise-containing images in line with actual conditions can be provided for a deep learning noise reduction task, and the robustness of a noise reduction model to identify different morphological noises is further improved.

[0079] In one possible embodiment, the noise generation module 102 is configured to: ​

[0080] determine a Poisson noise value of each pixel point in the noise-free image through a Poisson distribution formula;

[0081] generate a Poisson noise of the noise-free image based on the Poisson noise value of each pixel point.

[0082] In one possible embodiment, the noise generation module 102 is configured to:

[0083] determine a Poisson distribution expectation value of each pixel point in the noise-free image according to a pixel value of each pixel point;

[0084] calculate a Poisson noise value corresponding to the pixel value of each pixel point through a Poisson distribution formula and the Poisson distribution expectation value of each pixel point.

[0085] In one possible embodiment, the noise generation module 102 is configured to:

[0086] substitute the pixel value of each pixel point in the noise-free image into a Poisson mapping function relationship formula fitted by a polynomial to calculate the Poisson distribution expectation value of each pixel point.

[0087] In one possible embodiment, the noise generation module 102 is configured to:

[0088] determine a Gaussian noise value of each pixel point in the noise-free image through a Gaussian distribution formula, wherein a standard deviation of the Gaussian distribution formula is determined based on a distance of the pixel point to the center of the image, and an expectation value of the Gaussian distribution formula is 0;

[0089] generate an elliptical Gaussian noise of the noise-free image based on the Gaussian noise value of each pixel point.

[0090] In one possible embodiment, the noise generation module 102 is configured to:

[0091] determine a pixel distance value of each pixel point in the noise-free image to the center of the image, and determine a standard deviation of a Gaussian distribution corresponding to each pixel point according to the pixel distance value;

[0092] calculate a Gaussian noise value of each pixel point through a Gaussian distribution formula and the standard deviation of the Gaussian distribution corresponding to each pixel point.

[0093] In one possible embodiment, the noise generation module 102 is configured to:

[0094] substitute the pixel distance value into a Gaussian mapping function relationship formula fitted by a polynomial to calculate the standard deviation of the Gaussian distribution corresponding to each pixel point.

[0095] In a possible embodiment, the noise generation module 102 is configured to:

[0096] obtain a coefficient factor of the Poisson noise and a coefficient factor of the elliptical-like Gaussian noise;

[0097] multiply the Poisson noise and the elliptical-like Gaussian noise by the respective coefficient factors, and then superimpose the Poisson noise and the elliptical-like Gaussian noise into the noise-free image.

[0098] In a possible embodiment, the noise generation module 102 is further configured to:

[0099] obtain a first sample picture set before obtaining the noise-free image, the first sample picture set being composed of a plurality of frames of images continuously captured under the same shooting scene, wherein each frame of image contains one or more noise blocks;

[0100] obtain a second sample picture set, the second sample picture set being pictures captured by blocking the camera;

[0101] generate a Poisson noise generation model by calculating the first sample picture set, and generate an elliptical-like Gaussian noise generation model by calculating the second sample picture set.

[0102] In a possible embodiment, the noise generation module 102 is configured to:

[0103] calculate a pixel mean value and a pixel standard deviation of each pixel point contained in each noise block in each frame of image;

[0104] obtain a Poisson mapping function relationship formula of the pixel mean value to the pixel standard deviation by polynomial fitting.

[0105] In a possible embodiment, the noise generation module 102 is configured to:

[0106] determine an image center of each picture in the second sample picture set and a noise intensity value of each pixel point, and calculate a pixel distance value of each pixel point to the image center;

[0107] obtain a Gaussian mapping function relationship formula of the pixel distance value to the noise intensity value by polynomial fitting.

[0108] In a possible embodiment, the first sample picture set and the second sample picture set are respectively captured based on different device models and sensitivities, and different device models and sensitivities correspond to different Poisson noise generation models and elliptical-like Gaussian noise generation models.

[0109] Figure 8 A structural schematic diagram of a noise information generation device provided in an embodiment of the present application is shown in FIG. 1.Figure 8 As shown in the figure, the device includes a processor 201, a memory 202, an input device 203 and an output device 204; the number of processors 201 in the device can be one or more, Figure 8 In the embodiment, the processor 201 in the device, the memory 202, the input device 203 and the output device 204 can be connected through a bus or other means, Figure 8 In the embodiment, the connection through the bus is taken as an example. The memory 202 as a kind of computer readable storage medium, it can be used to store software program, computer executable program and module, such as the program instruction / module corresponding to the noise information generation method in the embodiment of the present application. The processor 201 executes the software program, instruction and module stored in the memory 202, thereby executing the various functional applications and data processing of the device, that is, realizing the noise information generation method described above. The input device 203 can be used to receive input digital or character information, and generate key signal input related to the user settings and function control of the device. The output device 204 can include display device such as display screen.

[0110] The embodiment of the present application also provides a storage medium containing computer executable instructions, the computer executable instructions are executed by computer processor to execute a noise information generation method described in the above embodiment, wherein, including:

[0111] Obtaining a noise-free image;

[0112] Determining the Poisson noise of the noise-free image through a Poisson noise generation model, and determining the class elliptical Gaussian noise of the noise-free image through a class elliptical Gaussian noise generation model, the Poisson noise generation model is used to generate noise related to image signal, and the class elliptical Gaussian noise generation model is used to generate noise irrelevant to image signal;

[0113] Superimposing the Poisson noise and the class elliptical Gaussian noise into the noise-free image to generate an image containing noise information.

[0114] It is worth noting that in the above embodiment of the noise information generation device, each unit and module included is only divided according to functional logic, but is not limited to the above division, as long as the corresponding function can be realized; in addition, the specific name of each functional unit is only for convenient mutual distinction, and does not limit the protection scope of the embodiment of the present application.

[0115] In some possible implementation, each of the aspects of the method provided by the present application can also be implemented in the form of a program product, which includes program codes for causing a computer device to perform the steps of the method according to various exemplary embodiments of the present application described above in the specification, for example, the computer device can perform the noise information generation method recorded in the embodiments of the present application. The program product can be implemented in any combination of one or more readable media.

Claims

1. A noise information generation method characterized by comprising: The method comprises the following steps: acquiring a noise-free image; determining Poisson noise of the noise-free image by a Poisson noise generation model and determining quasi-elliptical Gaussian noise of the noise-free image by a quasi-elliptical Gaussian noise generation model, the Poisson noise generation model being used to generate noise related to an image signal, the quasi-elliptical Gaussian noise generation model being used to generate noise irrelevant to the image signal, the quasi-elliptical Gaussian noise being Gaussian noise in the noise-free image varying with a distance value of each pixel point to a center of the image, and the noise intensity of the quasi-elliptical Gaussian noise being in an elliptical distribution gradually increasing from the center to the periphery; superimposing the Poisson noise and the quasi-elliptical Gaussian noise into the noise-free image to generate an image containing noise information.

2. The noise information generating method according to claim 1, characterized by, The step of determining the Poisson noise of the noise-free image by the Poisson noise generation model comprises the following steps: determining a Poisson noise value of each pixel point in the noise-free image by a Poisson distribution formula; generating the Poisson noise of the noise-free image based on the Poisson noise value of each pixel point.

3. The noise information generating method according to claim 2, characterized by, The step of determining the Poisson noise value of each pixel point in the noise-free image by the Poisson distribution formula comprises the following steps: determining a Poisson distribution expectation value of each pixel point according to a pixel value of each pixel point in the noise-free image; calculating a Poisson noise value corresponding to the pixel value of each pixel point by the Poisson distribution formula and the Poisson distribution expectation value of each pixel point.

4. The noise information generating method according to claim 3, characterized by, The step of determining the Poisson distribution expectation value of each pixel point according to the pixel value of each pixel point in the noise-free image comprises the following steps: substituting the pixel value of each pixel point in the noise-free image into a Poisson mapping function relationship obtained by polynomial fitting to calculate the Poisson distribution expectation value of each pixel point.

5. The noise information generation method according to claim 1, characterized by, The step of determining the quasi-elliptical Gaussian noise of the noise-free image by the quasi-elliptical Gaussian noise generation model comprises the following steps: determining a Gaussian noise value of each pixel point in the noise-free image by a Gaussian distribution formula, a standard deviation of the Gaussian distribution formula being determined based on a pixel distance value of each pixel point to a center of the image, and an expectation value of the Gaussian distribution formula being 0; generating the quasi-elliptical Gaussian noise of the noise-free image based on the Gaussian noise value of each pixel point.

6. The noise information generating method according to claim 5, wherein The step of determining the Gaussian noise value of each pixel point in the noise-free image by the Gaussian distribution formula comprises the following steps: determining a pixel distance value of each pixel point to the center of the image in the noise-free image, and determining a standard deviation of a Gaussian distribution corresponding to each pixel point according to the pixel distance value; calculating the Gaussian noise value of each pixel point by the Gaussian distribution formula and the standard deviation of the Gaussian distribution corresponding to each pixel point.

7. The noise information generation method according to claim 6, characterized by, The step of determining the standard deviation of the Gaussian distribution corresponding to each pixel point according to the pixel distance value comprises the following steps: substituting the pixel distance value into a Gaussian mapping function relationship obtained by polynomial fitting to calculate the standard deviation of the Gaussian distribution corresponding to each pixel point.

8. The noise information generation method according to any one of claims 1 to 5, characterized by, The step of superimposing the Poisson noise and the quasi-elliptical Gaussian noise into the noise-free image comprises the following steps: acquiring a coefficient factor of the Poisson noise and a coefficient factor of the quasi-elliptical Gaussian noise; The Poisson noise and the quasi-elliptical Gaussian noise are multiplied by respective coefficient factors and then superimposed into the noise-free image.

9. The noise information generation method according to claim 1, characterized by, Before the noise-free image is acquired, further comprising: acquiring a first sample picture set composed of multiple frames of images successively captured under the same shooting scene, wherein each frame of image contains one or more noise blocks; acquiring a second sample picture set composed of pictures captured by blocking the camera; calculating a Poisson noise generation model from the first sample picture set and a quasi-elliptical Gaussian noise generation model from the second sample picture set.

10. The noise information generating method according to claim 9, characterized by, The calculation of the Poisson noise generation model from the first sample picture set comprises: calculating the pixel mean and the pixel standard deviation of the pixel points contained in each noise block in each frame of image; calculating the Poisson mapping function relationship between the pixel mean and the pixel standard deviation by polynomial fitting.

11. The noise information generating method according to claim 9, wherein The calculation of the quasi-elliptical Gaussian noise generation model from the second sample picture set comprises: determining the image center of each picture in the second sample picture set and the noise intensity value of each pixel point, and calculating the pixel distance value of each pixel point to the image center; calculating the Gaussian mapping function relationship between the pixel distance value and the noise intensity value by polynomial fitting.

12. The noise information generating method according to any one of claims 9-11, characterized by, The first sample picture set and the second sample picture set are respectively captured based on different device models and sensitivities, and different device models and sensitivities correspond to different Poisson noise generation models and quasi-elliptical Gaussian noise generation models.

13. A noise information generating apparatus characterized by comprising: comprising: an image acquisition module configured to acquire a noise-free image; a noise generation module configured to determine the Poisson noise of the noise-free image by a Poisson noise generation model and the quasi-elliptical Gaussian noise of the noise-free image by a quasi-elliptical Gaussian noise generation model, the Poisson noise generation model being used to generate noise related to image signal, the quasi-elliptical Gaussian noise generation model being used to generate noise irrelevant to image signal, the quasi-elliptical Gaussian noise being Gaussian noise varying with the distance value of each pixel point to the image center in the noise-free image, and its noise intensity being in an elliptical distribution gradually increasing from the center to the periphery; a noise addition module configured to superimpose the Poisson noise and the quasi-elliptical Gaussian noise into the noise-free image to generate an image containing noise information.

14. A noise information generating apparatus, the apparatus comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the noise information generation method of any one of claims 1-12.

15. A storage medium storing computer executable instructions for executing the noise information generation method of any one of claims 1-12 when executed by a computer processor.

16. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the noise information generation method of any one of claims 1-12.

Images