Video denoising data set generation method, computer device and computer readable storage medium

Abstract

The invention provides a video denoising data set generation method, a computer device and a computer readable storage medium, and the method comprises the steps: shooting a clean-noise image pair needed by training, and shooting a clean video needed by generating a noise video; inputting the clean image into a noise adding neural network, carrying out difference calculation on a training image output by the noise adding neural network and the noise image to obtain a first difference value, reversely inputting the first difference value into the noise adding neural network, and updating parameters by the noise adding neural network according to the first difference value; after the parameters are updated for multiple times, a trained target noise adding neural network is obtained; and inputting the clean video into the target noise adding neural network to generate a noise video, wherein the noise video is an RAW video. The invention further provides a computer device and a computer readable storage medium for implementing the method. According to the invention, the quality of the data set can be improved, and the generated noise video is closer to a real shot video.

Description

Technical Field

[0001] The present invention relates to the technical field of video denoising data training, and in particular, to a method for generating a video denoising data set, as well as a computer device and a computer-readable storage medium for implementing the method. Background Art

[0002] In low light conditions, the original images obtained by the image sensor often contain a lot of noise. In order to improve the quality of the output image, it is usually necessary to denoise the image. After denoising the image with large noise, the image's perception will be significantly improved. In addition, image denoising has different benefits in different application scenarios. For example, the video after image denoising can reduce the bit rate of video encoding.

[0003] At present, common image denoising methods are divided into two categories, namely traditional algorithms and algorithms based on deep learning. The denoising effect of current deep learning algorithms is better than that of traditional algorithms, and has become a hot topic of research. The RAW images obtained by the image sensor contain more original information, and the distribution of noise is not destroyed by the subsequent processing flow, which is more conducive to denoising. Therefore, the clean video and noisy RAW video data set is the basis for training deep learning denoising algorithms. Generally, the data set is made by real shooting or synthesis. The real shooting method is to shoot still images by setting the same exposure but different exposure time and gain combination. The dynamic scene is simulated by stop-motion animation or shooting the screen content, but the stop-motion animation cannot restore the characteristics of the real dynamic scene, such as motion blur, deformation of objects, and real dynamics of moving objects; and the screen that pauses and shoots the motion scene frame by frame will have other problems, such as the screen spectrum is inconsistent with the spectrum of the real scene, the pixel grid is captured, the screen dynamic range is limited, and the screen pixels may be lower than the image sensor pixels. The synthesis method can solve all the problems of real shooting. The general synthesis method is to add noise to the clean RAW video frame by frame. The clean RAW video is the RAW video shot by the camera at the lowest gain. The way of adding noise directly determines the quality of the dataset, and thus the effect of the final denoising algorithm.

[0004] Therefore, image denoising based on deep learning algorithms needs to provide a method for applying clean RAW images to generate corresponding noisy RAW images, and the generated noisy RAW images should be as close as possible to the actual images taken at the same time and in the same scene as the clean RAW images when the image sensor is set to a specific high gain. At present, the commonly used method is to first establish a noise model, determine the model parameters by calibration after the noise model is established, and use the noise model to add noise to the clean RAW image to generate the corresponding noisy RAW image. The commonly used noise model is the Poisson-Gaussian mixed noise model. For example, the solution with publication number CN111724317A discloses a method for constructing a RAW image denoising dataset using a Poisson-Gaussian mixed noise model.

[0005] However, most of these generative solutions use the Poisson-Gaussian mixed noise model, which cannot fully describe the actual noise, resulting in differences between the generated noise image and the actual image. In addition, the existing process of generating noise images is not processed entirely in the RAW domain, but uses sRGB images to generate RAW images through various conversion methods and then processes the RAW images, introducing more variables and amplifying the difference between the generated noise images and the actual images.

[0006] For example, the invention patent application with publication number CN116563755A discloses a method for generating a real-world RAW video denoising dataset based on screen capture images. This is a screen capture-based generation method, but because the screen spectrum is inconsistent with the real scene spectrum, and the pixel grid and screen dynamic range are limited, the screen pixels may be lower than the image sensor pixels, resulting in differences between the generated noise image and the real image, and ultimately the trained deep learning algorithm cannot perform denoising on the noise image well. Summary of the invention

[0007] The first object of the present invention is to provide a method for generating a video denoising data set in which the obtained noisy video is closer to the real shot video.

[0008] The second objective of the present invention is to provide a computer device for implementing the above-mentioned video denoising data set generation method.

[0009] A third object of the present invention is to provide a computer-readable storage medium for implementing the above-mentioned video denoising dataset generation method.

[0010] To achieve the first purpose of the present invention, the video denoising data set generation method provided by the present invention includes shooting a clean-noise image pair required for training, the clean-noise image pair includes a clean image and a noisy image, and shooting a clean video required for generating a noisy video, the clean image and the noisy image are both RAW images, and the clean video is a RAW video; and, training a noise adding neural network for adding noise: inputting the clean image into the noise adding neural network, performing a difference calculation between the training image output by the noise adding neural network and the noise image to obtain a first difference, reversely inputting the first difference into the noise adding neural network, the noise adding neural network updates parameters according to the first difference, and outputs a new training image again; after updating the parameters for multiple times, stopping updating the parameters after satisfying a preset condition for stopping updating the parameters, and obtaining a trained target noise adding neural network; inputting the clean video into the target noise adding neural network to generate a noisy video, and the noisy video is a RAW video.

[0011] It can be seen from the above scheme that in the process of generating the noise image, the whole process is processed in the RAW domain, that is, the clean image and the noise image obtained are all RAW images, and the clean video is also a RAW video. The images used in the noise adding neural network training process are all RAW images. Therefore, it can avoid the problem of image information loss caused by image format conversion, which leads to a large difference between the generated noise image and the real shot image, and can improve the quality of the noise video finally obtained. In addition, since the data set is not generated using the screen capture method, the defects of the screen capture method can be avoided.

[0012] A preferred solution is that when training the noise adding neural network, the training image output by the noise adding neural network is also output to a preset real-shot image judgment neural network. The real-shot image judgment neural network also receives the noise image, and a difference calculation is performed on the training image output by the noise adding neural network and the noise image to obtain a second difference. The second difference is added to the first difference and then reversely input into the noise adding neural network.

[0013] It can be seen that by further calculating the second difference between the training image and the noise image through an additional real-shot image judgment neural network, and adding the second difference to the first difference and then outputting it to the noise adding neural network, the accuracy of the noise adding neural network parameter adjustment can be improved.

[0014] A further solution is that adding the second difference to the first difference and then inputting the resultant to the noise adding neural network includes: performing weighted calculation on the first difference and the second difference and then outputting the calculation result to the noise adding neural network.

[0015] In this way, by setting appropriate weighting values, the accuracy of noise addition neural network parameter adjustment can be further improved.

[0016] A further solution is that the preset condition for stopping updating the parameters includes: a first difference calculated by taking a difference between a training image output by the noise-added neural network and a noise image satisfies a preset condition.

[0017] An optional solution is that the preset condition for stopping updating the parameters includes: the number of times the parameters are updated reaches a preset upper limit.

[0018] It can be seen that the parameter update will stop after certain conditions are met, ensuring the calculation accuracy while taking into account the calculation efficiency.

[0019] A further solution is that the number of the clean-noise image pairs is multiple, and the gain of the noise image of at least one pair of the clean-noise image pairs is different from the gain of the noise image of another pair of the clean-noise image pairs.

[0020] It can be seen that by setting a pair of clean-noise images with different increases and thereby adding noise to the neural network, multiple target noise-adding neural networks with different gains can be obtained.

[0021] A further solution is to use clean-noise image pairs with different gains to train multiple target noise adding neural networks for different gains.

[0022] In this way, the gains corresponding to the multiple target noise adding neural networks are different. When the noisy video is subsequently acquired, the appropriate target noise adding neural network can be selected as needed to generate the corresponding noisy video.

[0023] A further solution is that the clean video is input into a target noise adding neural network to generate a noisy video, including: inputting a clean video into multiple target noise adding neural networks with different gains to obtain multiple noisy videos with different gains.

[0024] It can be seen that based on adding neural networks with different added target noises, different noisy videos can be generated for the same clean video, and different noisy videos have different gains, which makes the data set richer and is conducive to the training of the denoising neural network.

[0025] To achieve the above second objective, the computer device provided by the present invention includes a processor and a memory, the memory stores a computer program, and when the computer program is executed by the processor, each step of the above video denoising data set generation method is implemented.

[0026] To achieve the third objective mentioned above, the present invention provides a computer readable storage medium storing a computer program, and when the computer program is executed by a processor, each step of the above-mentioned method for generating a video denoising data set is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of an embodiment of a method for generating a video denoising data set according to the present invention.

[0028] Figure 2 Schematic diagram of a clean-noise image pair obtained by an embodiment of the method for generating a video denoising dataset of the present invention.

[0029] Figure 3 Schematic diagram of a clean video obtained by an embodiment of the method for generating a video denoising dataset of the present invention.

[0030] Figure 4 It is a flowchart of the first method of training a noise adding neural network in the embodiment of the method for generating a video denoising data set of the present invention.

[0031] Figure 5 It is a flowchart of the second method of training a noise adding neural network in the embodiment of the video denoising data set generation method of the present invention.

[0032] Figure 6 It is a flowchart of generating a noisy video according to an embodiment of the video denoising data set generation method of the present invention.

[0033] Figure 7 It is a schematic diagram of a noisy video obtained by an embodiment of the method for generating a video denoising data set of the present invention.

[0034] The present invention is further described below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0035] The video denoising data set generation method of the present invention is used to generate a data set used by a denoising deep learning algorithm, and these data sets include clean videos and noisy videos with noise, and the noisy videos are generated by adding noise to the clean videos. The video denoising data set generation method of the present invention can be implemented on a computer device, and the computer device is provided with a processor and a memory, and a computer program is stored on the memory. The processor implements the above-mentioned video denoising data set generation method by executing the computer program.

[0036] Video denoising dataset generation method embodiment:

[0037] This embodiment mainly includes three major steps. The first step is to shoot the clean-noise image pairs required for training and the clean video required to generate the noisy video. The second step is to train the noise adding neural network. The third step is to use the trained noise adding neural network to add noise to the clean video. The shot clean video and the noisy video after adding noise form a clean-noisy video pair, and multiple clean-noisy video pairs are video denoising data sets. The whole process of the above three major steps is implemented in the RAW domain, that is, the image and video are not converted in format, and the original information of the image can be retained as much as possible, so that the obtained noisy video is closer to the real shot video.

[0038] See also Figure 1 , this embodiment first executes step S11 to shoot the clean-noise image pair required for training, wherein the clean-noise image pair includes a clean image and a noisy image, and both the clean image and the noisy image are RAW images, that is, the images originally obtained by the image sensor. Specifically, when shooting a clean image, the image sensor is set to the lowest gain, and the image is captured at the lowest gain, and when shooting a noisy image, the image sensor is set to be greater than the lowest gain, and the image is captured under a higher gain condition. When shooting the same pair of clean-noise image pairs, it is necessary to keep the camera and the captured image completely still, and the settings for shooting clean images and noisy images are only different in gain and exposure time, and the exposure of the two images of the same pair of clean-noise image pairs is the same, that is, the product of gain and exposure time is the same. Preferably, the number of clean-noise image pairs shot can be multiple pairs, and the clean-noise image pairs shot should contain as many scenes as possible and cover as many gain values ​​below the required maximum gain as possible. As Figure 2 As shown, under the first gain value, that is, under the condition that the gain value of the camera is set to "gain 1", multiple pairs of clean-noise image pairs can be obtained, wherein the first pair of clean-noise image pairs includes the lowest gain RAW image 1-1 and the corresponding high gain RAW image 1-1, the second pair of clean-noise image pairs includes the lowest gain RAW image 1-2 and the corresponding high gain RAW image 1-2, and so on. Under the condition that the gain value is set to "gain 2", multiple pairs of clean-noise image pairs are also obtained, wherein the first pair of clean-noise image pairs includes the lowest gain RAW image 2-1 and the corresponding high gain RAW image 2-1, the second pair of clean-noise image pairs includes the lowest gain RAW image 2-2 and the corresponding high gain RAW image 2-2, and so on. In this way, multiple pairs of clean-noise image pairs can be obtained for multiple different gain values.

[0039] In addition, shoot a clean video required to generate the noisy video. The clean video is also a RAW video. When shooting a clean video, set the image sensor to the lowest gain, the frame rate to the commonly used frame rate, and the exposure time to the longest exposure time or close to the longest exposure time under the corresponding frame rate. The exposure can be controlled by changing the aperture or using a neutral density filter. The captured video screen does not have to be completely still, it can contain global displacement or local displacement, and should be as close to commonly used scenes as possible and contain as many different lighting conditions as possible. Figure 3 As shown, for different scenes or different lighting conditions, multiple clean videos are obtained respectively. In fact, each clean video is a minimum gain RAW video, and each clean video contains multiple RAW images captured under the minimum gain condition. Therefore, each clean video is also a set of clean image sequences.

[0040] Then, the noise adding neural network for adding noise is trained. The first training method for the noise adding neural network is to input a clean image into the noise adding neural network, and the noise adding neural network performs noise adding processing on the clean image based on the current parameters to obtain a training image, that is, to execute step S12. In the initial state, the initial noise adding neural network needs to be pre-set, and each parameter in the initial noise adding neural network is an initial value.

[0041] For details, see Figure 4 , the lowest gain RAW image 31, i.e., the clean image, is input into the noise adding neural network 32, and the noise adding neural network 32 performs noise adding processing on the clean image to obtain a training image 33, wherein the training image 33 is also a RAW image. Then, step S13 is executed, the training image 33 and the noise image 34 in the clean-noise image pair are input into the loss function 35, and the loss function 35 is used to perform a difference calculation on the training image 35 and the noise image 34 to obtain a first difference, and the first difference is reversely input into the noise adding neural network 32.

[0042] Next, step S14 is executed, and the noise adding neural network 32 updates its own parameters according to the first difference, so that the output training image is closer to the noise image 34. In this way, the noise adding neural network 32 with updated parameters performs noise adding processing on the clean image again, obtains a new training image 33 again, obtains the training image 33 again, inputs the training image 33 into the loss function 35 again, and performs difference calculation with the noise image 34 again.

[0043] At this time, step S15 needs to be executed to determine whether the condition for stopping updating the parameters is met. Preferably, the condition for updating the parameters can be: the first difference between the training image and the noise image output by the noise addition neural network 32 meets the preset requirements, for example, the first difference is less than the preset difference threshold; or, the number of times the parameters are updated reaches the preset upper limit, for example, the upper limit can be 10 times or 20 times. If any of the above two conditions is met, it is considered that the condition for stopping updating the parameters is met, the operation of updating the parameters is stopped, and step S16 is executed. If the condition for updating the parameters is not met, return to step S12 and continue to perform the next round of calculations.

[0044] After multiple parameter update calculations, a trained noise adding neural network is obtained, which is called the target noise adding neural network.

[0045] The second way to train a neural network for noise addition is to see Figure 5 , the lowest gain RAW image 41, i.e., the clean image, is input into the noise adding neural network 42, and the noise adding neural network 42 performs noise adding processing on the clean image to obtain a training image 43, wherein the training image 43 is also a RAW image. Then, the training image 43 and the noise image 44 in the clean-noise image pair are input into the loss function 45, and the loss function 45 is used to perform a difference calculation on the training image 45 and the noise image 44 to obtain a first difference.

[0046] Compared with the first training method, the second training method also inputs the training image 43 and the noise image 44 into a real-shot image judgment neural network 46. Preferably, the real-shot image judgment neural network 46 is a pre-set neural network, and is used to calculate the difference between two images to determine whether the image is a real-shot image. The real-shot image judgment neural network 46 can be implemented using a known neural network.

[0047] The real-shot image judgment neural network 46 obtains a second difference after performing a difference calculation on the training image 43 and the noise image 44, and adds the second difference to the first difference and then inputs it back to the noise adding neural network 42. Preferably, when the second difference is added to the first difference, the first difference and the second difference are weighted, for example, the weight of the first difference and the weight of the second difference are set, and a weighted sum calculation is performed, and the calculation result is input back to the noise adding neural network 42.

[0048] Next, the noise adding neural network 42 updates its own parameters according to the input difference, so that the output training image is closer to the noise image 44. In this way, the noise adding neural network 42 with updated parameters performs noise adding processing on the clean image again, obtains a new training image 43 again, obtains the training image 43 again, inputs the training image 43 into the loss function 45 again and calculates the difference with the noise image 44 again, and so on, for multiple rounds of calculations.

[0049] After updating the parameters for multiple times, the updating of the parameters is stopped when the preset condition for stopping the updating of the parameters is met, and the trained target noise adding neural network is obtained, that is, step S16 is executed.

[0050] Finally, step S17 is performed to input the clean video into the target noise adding neural network to generate a noisy video, and the generated noisy video is also a RAW video. Figure 6 , the minimum gain RAW video 51 is input into the target noise adding neural network 52. In fact, the minimum gain RAW video 51 is a sequence of multiple images. The target noise adding neural network 52 performs noise adding calculation on each image of the image sequence to obtain a noise image sequence, thus forming a noise video 53.

[0051] Preferably, Figure 2 As shown, when generating clean-noise image pairs, multiple pairs of clean-noise image pairs are generated for different gains. In this way, when training and generating the target noise adding neural network, corresponding target noise adding neural networks can be generated for different gains. In other words, there are multiple target noise adding neural networks generated, and each target noise adding neural network corresponds to a different gain value. In this way, in step S17, for the same clean video, multiple noise videos can be generated using target noise adding neural networks with different gain values, and each noise video corresponds to a different gain value. Figure 7 As shown, for example, the clean video is the lowest gain RAW image sequence 4-1, and for different gain values, a high gain RAW image sequence 5-1 and a high gain RAW image sequence 6-1 are generated respectively, and even a high gain RAW image sequence for other gain values ​​can be generated. Correspondingly, for another clean video, for example, the clean video is the lowest gain RAW image sequence 4-2, and for different gain values, a high gain RAW image sequence 5-2 and a high gain RAW image sequence 6-2 are generated respectively, and so on.

[0052] In this way, a data set containing clean videos and noisy videos is finally generated. The data set contains multiple video pairs, and a video pair includes a clean video and at least one noisy video. If multiple noisy videos are included, the gain corresponding to each noisy video is different.

[0053] It can be seen that the entire process of training the target noise addition neural network in this embodiment is carried out in the RAW domain. Therefore, the generated noisy video is very similar to the actual video. Therefore, the obtained data set can be closer to the actual video. Training the deep learning algorithm with the data set formed in this embodiment can enhance the denoising ability of the deep learning algorithm and make the denoising performance of the deep learning algorithm better.

[0054] Computer device embodiment:

[0055] The computer device of this embodiment can be an electronic device with human-computer interaction function, such as a smart terminal device. The electronic device has a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an information processing program for implementing the above-mentioned information processing method. When the processor executes the computer program, each step of the above-mentioned video denoising data set generation method is implemented.

[0056] For example, a computer program may be divided into one or more modules, one or more modules are stored in a memory and executed by a processor to complete each module of the present invention. One or more modules may be a series of computer program instruction segments capable of completing a specific function, and the instruction segments are used to describe the execution process of the computer program in a terminal device.

[0057] It should be noted that the terminal device may be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the schematic diagram of the present invention is only an example of a terminal device and does not constitute a limitation on the terminal device. The terminal device may include more or fewer components than shown in the figure, or may combine certain components, or different components. For example, the terminal device may also include an input / output device, a network access device, a bus, etc.

[0058] The processor referred to in the present invention may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0059] The memory can be used to store computer programs and / or modules. The processor realizes various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0060] Computer readable storage medium:

[0061] If the computer program stored in the computer device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, each step of the above-mentioned video denoising data set generation method can be implemented.

[0062] Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form, etc. Computer readable media may include: any entity or device capable of carrying computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer readable media do not include electric carrier signals and telecommunication signals.

[0063] Finally, it should be emphasized that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Video denoising dataset generation method, including: Shooting a clean-noise image pair required for training, wherein the clean-noise image pair includes a clean image and a noisy image, and shooting a clean video required for generating a noisy video, wherein both the clean image and the noisy image are RAW images, and the clean video is a RAW video; Features: Training a noise adding neural network for adding noise: inputting the clean image into the noise adding neural network, performing a difference calculation between a training image output by the noise adding neural network and the noise image to obtain a first difference, reversely inputting the first difference into the noise adding neural network, the noise adding neural network updates parameters according to the first difference, and outputs a new training image again; After updating the parameters for multiple times, the parameter updating is stopped when the preset condition for stopping the parameter updating is met, and the trained target noise adding neural network is obtained; The clean video is input into the target noise adding neural network to generate the noisy video, where the noisy video is a RAW video.

2. The method for generating a video denoising dataset according to claim 1, characterized in that: When training the noise adding neural network, the training image output by the noise adding neural network is also output to a preset real-shot image judgment neural network. The real-shot image judgment neural network also receives the noise image, performs a difference calculation on the training image output by the noise adding neural network and the noise image to obtain a second difference, adds the second difference to the first difference, and then inputs the second difference back into the noise adding neural network.

3. The method for generating a video denoising dataset according to claim 2, characterized in that: Adding the second difference to the first difference and then inputting the second difference to the noise adding neural network in reverse order comprises: After weighted calculation is performed on the first difference and the second difference, the calculation result is output to the noise adding neural network.

4. The method for generating a video denoising dataset according to any one of claims 1 to 3, characterized in that: The preset condition for stopping updating parameters includes: taking a difference between the training image output by the noise addition neural network and the noise image, and the first difference calculated satisfies a preset condition.

5. The method for generating a video denoising dataset according to any one of claims 1 to 3, characterized in that: The preset condition for stopping updating the parameters includes: the number of times the parameters are updated reaches a preset upper limit.

6. The method for generating a video denoising dataset according to any one of claims 1 to 3, characterized in that: The number of the clean-noise image pairs is plural, and the gain of the noise image of at least one pair of the clean-noise image pairs is different from the gain of the noise image of another pair of the clean-noise image pairs.

7. The method for generating a video denoising dataset according to claim 6, characterized in that: A plurality of target noise adding neural networks with different gains are obtained by training clean-noise image pairs with different gains.

8. The method for generating a video denoising dataset according to claim 7, characterized in that: Inputting the clean video into the target noise adding neural network to generate the noisy video comprises: A clean video is input into a plurality of target noise adding neural networks with different gains to obtain a plurality of noisy videos with different gains.

9. A computer device, characterized in that The method comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, each step of the method for generating a video denoising data set as claimed in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for generating a video denoising data set as described in any one of claims 1 to 8 is implemented.

Images