CN-121985080-A - Method and system for constructing video noise reduction data set and image processor

Abstract

The invention provides a method, a system and an image processor for constructing a video noise reduction data set, wherein the method comprises the steps of carrying out data expansion based on a truly photographed video data set to obtain a first data set; the method comprises the steps of obtaining a first data set by adding first noise and performing first clipping processing based on a dynamic video noise reduction data set, obtaining a third data set by adding second noise and performing second clipping processing based on a static video noise reduction data set, and constructing a video noise reduction data set based on the first data set, the second data set and the third data set. The method, the system and the image processor for constructing the video noise reduction data set reduce the construction cost of the video noise reduction data set and improve the construction efficiency of the video noise reduction data set.

Inventors

• LI ZHENG
• FENG JINLI
• SHEN CHAO

Assignees

• 北京清微智能科技股份有限公司

Dates

Publication Date

20260505

Application Date

20260109

Claims (17)

1. 1. A method of constructing a video noise reduction dataset, comprising: Performing data expansion based on a truly photographed video data set to obtain a first data set; Performing first noise addition and first clipping processing based on the dynamic video noise reduction data set to obtain a second data set; Performing second noise addition and second clipping processing based on the static video noise reduction data set to obtain a third data set; a video noise reduction dataset is constructed based on the first dataset, the second dataset, and the third dataset.
2. 2. The method of claim 1, wherein the data expansion based on the truly captured video data set to obtain the first data set comprises: The method comprises the steps of obtaining a real shot video data set, wherein the real shot video data set comprises black frame video stream data, 24-color card video stream data and indoor static scene video stream data under different light intensities shot by an image sensor; and cutting out the fixed size based on the truly photographed video data set to obtain the first data set.
3. 3. The method of claim 2, wherein the cropping of a fixed size based on the truly captured video data set to obtain the first data set comprises: and cutting each frame of image in the truly shot video data set according to a fixed size sequentially from a preset position or randomly selecting a starting position to cut according to the fixed size, so as to obtain a multi-frame cutting image corresponding to each frame of image.
4. 4. The method of claim 1, wherein performing a first noise addition based on the dynamic video noise reduction dataset comprises: performing black frame sampling on the first intermediate data set to obtain illumination independent noise corresponding to each frame image in the dynamic video noise reduction data set; And adding noise to each frame of image in the dynamic video noise reduction data set according to the noise which corresponds to each frame of image in the dynamic video noise reduction data set and is irrelevant to illumination, so as to obtain a noise dynamic data set, wherein the noise which is relevant to illumination is obtained in advance.
5. 5. The method of claim 4, wherein the noise adding each frame of the images in the dynamic video noise reduction dataset according to the illumination independent noise corresponding to each frame of the images in the dynamic video noise reduction dataset and the illumination dependent noise comprises: And adding noise to each frame of image in the dynamic video noise reduction data set according to a formula D 1 = KI 1 + N 1 , wherein D 1 represents one frame of image in the noise dynamic data set, I 1 represents one frame of image in the dynamic video noise reduction data set, N 1 represents illumination-independent noise corresponding to one frame of image in the dynamic video noise reduction data set, K represents illumination-related noise, and noise calibration is carried out on the basis of one frame of image of 24-color card video stream data included in the truly shot video stream data.
6. 6. The method of claim 4, wherein performing a compensated black frame quantization truncation process on the dynamic video noise reduction dataset to obtain a first intermediate dataset comprises: Screening the dynamic video noise reduction data set based on a first sensitivity threshold to obtain a temporary dynamic video data set; drawing a brightness histogram of each frame of image in the temporary dynamic video data set to obtain a brightness histogram corresponding to each frame of image in the temporary dynamic video data set; Performing truncation processing on a brightness histogram corresponding to each frame of image in the temporary dynamic video data set to obtain a truncated brightness histogram corresponding to each frame of image in the temporary dynamic video data set; and compensating the truncated brightness histogram corresponding to each frame of image in the temporary dynamic video data set, so that the compensated brightness histogram corresponding to each frame of image in the temporary dynamic video data set accords with Gaussian distribution.
7. 7. The method of claim 5, wherein performing a first clipping process based on the noise dynamic data set comprises: and cutting the motion area of the continuous preset number of frame images of the noise dynamic data set.
8. 8. The method of claim 1, wherein performing a second noise addition based on the still video noise reduction dataset comprises: performing compensation black frame quantization truncation processing on the static video noise reduction data set to obtain a second intermediate data set; black frame sampling is carried out on the second intermediate data set, and noise which corresponds to each frame of image and is irrelevant to illumination in the static video noise reduction data set is obtained; And adding noise to each frame of image in the static video noise reduction data set according to the noise which corresponds to each frame of image in the static video noise reduction data set and is irrelevant to illumination, so as to obtain a noise static data set, wherein the noise which is relevant to illumination is obtained in advance.
9. 9. The method of claim 8, wherein the noise adding each frame of the still video noise reduction dataset based on the illumination independent noise corresponding to each frame of the image in the still video noise reduction dataset and the illumination dependent noise comprises: And adding noise to each frame of image in the static video noise reduction data according to a formula D 2 = KI 2 + N 2 , wherein D 2 represents one frame of image in the noise static data set, I 2 represents one frame of image in the static video noise reduction data, N 2 represents illumination-independent noise corresponding to one frame of image in the static video noise reduction data, K represents illumination-related noise, and noise calibration is carried out on the basis of one frame of image of 24-color card video stream data included in the truly shot video stream data.
10. 10. The method of claim 8, wherein performing a compensated black frame quantization truncation process on the still video noise reduction dataset to obtain a second intermediate dataset comprises: screening the static video noise reduction data set based on a second sensitivity threshold to obtain a temporary static video data set; drawing a brightness histogram of each frame of image in the temporary static video data set to obtain a brightness histogram corresponding to each frame of image in the temporary static video data set; performing truncation processing on a brightness histogram corresponding to each frame of image in the temporary static video data set to obtain a truncated brightness histogram corresponding to each frame of image in the temporary static video data set; And compensating the truncated brightness histogram corresponding to each frame of image in the temporary static video data set, so that the compensated brightness histogram corresponding to each frame of image in the temporary static video data set accords with Gaussian distribution.
11. 11. The method of claim 8, wherein performing a second clipping process based on the noisy static data set comprises: And cutting each frame of image in the noise static data set according to a fixed size sequentially from a preset position, randomly selecting a starting position to cut according to the fixed size, or randomly moving a preset distance by the random selection starting position to cut according to the fixed size, so as to obtain a multi-frame cutting image corresponding to each frame of image.
12. 12. A training system for a noise reduction network model, characterized in that the video noise reduction data set obtained by the method for constructing a video noise reduction data set according to claims 1 to 11 comprises: the acquisition module is used for acquiring a video noise reduction data set as training data; and the training module is used for training the original model based on the training data to obtain a noise reduction network model.
13. 13. An image processor, characterized in that the noise reduction network model obtained by the training system for noise reduction network model according to claim 12 comprises: The receiving module is used for receiving the video stream data to be noise reduced; And the noise reduction module is used for carrying out noise reduction processing on the video stream data to be noise reduced through the noise reduction network model.
14. 14. A board comprising at least one image processor according to claim 13.
15. 15. An electronic device comprising at least one board as claimed in claim 14.
16. 16. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the method of any one of claims 1 to 11.
17. 17. A computer program product comprising computer programs/instructions which, when executed by a processor, implement the method of any one of claims 1 to 11.

Description

Method and system for constructing video noise reduction data set and image processor Technical Field The invention relates to the technical field of image processing, in particular to a method and a system for constructing a video noise reduction data set and an image processor. Background The image signal processing is to process the original data output by the image sensor to make the original data accord with the real feeling of human eyes. Devices such as smart phones and security monitoring terminals which can shoot images all need to process image signals so as to improve image quality. In order to perform noise reduction processing on video data, a deep learning training model is generally adopted to obtain a noise reduction model. In performing noise reduction model training, a large amount of training data is required. In the prior art, in order to obtain a high-quality clean-noisy original data (Raw) video data pair, a real noisy-clean Raw video frame data pair can be acquired based on a screen shot image, and related data processing is performed to obtain a Raw video denoising data set. The acquiring process of the Raw video denoising data set is highly dependent on high-quality display and shooting equipment, needs to shoot a large amount of video data, consumes a large amount of time and has high cost. Therefore, how to reduce the construction cost of the video noise reduction dataset is an important issue to be solved in the art. Disclosure of Invention Aiming at the problems in the prior art, the embodiment of the invention provides a method and a system for constructing a video noise reduction data set and an image processor, which can at least partially solve the problems in the prior art. In a first aspect, the present invention proposes a method for constructing a video noise reduction dataset, comprising: Performing data expansion based on a truly photographed video data set to obtain a first data set; Performing first noise addition and first clipping processing based on the dynamic video noise reduction data set to obtain a second data set; Performing second noise addition and second clipping processing based on the static video noise reduction data set to obtain a third data set; a video noise reduction dataset is constructed based on the first dataset, the second dataset, and the third dataset. Further, the performing data expansion based on the truly photographed video data set, and obtaining the first data set includes: The method comprises the steps of obtaining a real shot video data set, wherein the real shot video data set comprises black frame video stream data, 24-color card video stream data and indoor static scene video stream data under different light intensities shot by an image sensor; and cutting out the fixed size based on the truly photographed video data set to obtain the first data set. Further, the performing fixed-size cropping on the first intermediate data set, and obtaining the first data set includes: and cutting each frame of image in the truly shot video data set according to a fixed size sequentially from a preset position or randomly selecting a starting position to cut according to the fixed size, so as to obtain a multi-frame cutting image corresponding to each frame of image. Further, performing a first noise addition to the dynamic video noise reduction dataset includes: performing black frame sampling on the first intermediate data set to obtain illumination independent noise corresponding to each frame image in the dynamic video noise reduction data set; And adding noise to each frame of image in the dynamic video noise reduction data set according to the noise which corresponds to each frame of image in the dynamic video noise reduction data set and is irrelevant to illumination, so as to obtain a noise dynamic data set, wherein the noise which is relevant to illumination is obtained in advance. Further, the noise adding to each frame of image in the dynamic video noise reduction dataset according to the noise which corresponds to each frame of image in the dynamic video noise reduction dataset and is irrelevant to illumination, and the obtaining of the noise dynamic dataset includes: And adding noise to each frame of image in the dynamic video noise reduction data set according to a formula D 1 = KI1 + N1, wherein D 1 represents one frame of image in the noise dynamic data set, I 1 represents one frame of image in the dynamic video noise reduction data set, N 1 represents illumination-independent noise corresponding to one frame of image in the dynamic video noise reduction data set, K represents illumination-related noise, and noise calibration is carried out on the basis of one frame of image of 24-color card video stream data included in the truly shot video stream data. Further, the performing black frame quantization truncation processing on the dynamic video noise reduction data set to obtain a first intermediate data set includes: Screenin