Image denoising model training method, imaging denoising method, devices and storage medium

1. A training method for an image denoising model, comprising:

collecting multiple sample image groups through a shooting device, each of the sample image groups including multiple frames of sample images with a same photographic sensitivity, and sample images in different sample image groups have different photographic sensitivities;

for each of the sample image groups,

acquiring a photographic sensitivity of the sample image group;

constructing a photographic sensitivity image based on the photographic sensitivity and determining a noise characterization image corresponding to the sample image group based on the photographic sensitivity image, wherein all pixel values of the photographic sensitivity image are a normalized photographic sensitivity of the sample image group; and

determining a training input image group and a target image which are associated with the sample image group, the training input image group including all or part of the sample images in the sample image group and the noise characterization image corresponding to the sample image group;

constructing multiple training pairs, each of the training pairs including a training input image group and a corresponding target image; and

training the image denoising model based on the multiple training pairs until the image denoising model converges.

2. The training method of claim 1 , wherein for each of the sample image groups, determining the noise characterization image corresponding to the sample image group based on the photographic sensitivity image further comprises:

determining the photographic sensitivity image as the noise characterization image.

3. The training method of claim 1 , wherein for each of the sample image groups, determining the noise characterization image corresponding to the sample image group based on the photographic sensitivity image further comprises:

for each of the sample image groups, acquiring a grayscale component image of the target image corresponding to the sample image group, determining a dot product matrix of the grayscale component image and the photographic sensitivity image, and determining the dot product matrix as the noise characterization image.

4. The training method of claim 1 , wherein for each of the sample image groups, determining the noise characterization image corresponding to the sample image group based on the photographic sensitivity image further comprises:

for each of the sample image groups, determining a grayscale component image of the target image corresponding to the sample image group; and

determining that the noise characterization image is a two-dimensional image having the photographic sensitivity image and the grayscale component image.

5. The training method of claim 1 , wherein the normalized photographic sensitivity is obtained by:

making statistics on a maximum photographic sensitivity and a minimum photographic sensitivity in the multiple sample image groups, determining a photographic sensitivity interval based on the maximum photographic sensitivity and the minimum photographic sensitivity, constructing a mapping relationship between the photographic sensitivity interval and an image pixel interval, and for each of the sample image groups, determining the normalized photographic sensitivity of the sample image group based on the mapping relationship.

6. The training method of claim 1 , wherein for each of the sample image groups, determining the target image associated with the sample image group comprises one of following operations:

for each of the sample image groups,

calculating an average of pixels at a same position based on a sample image in the sample image group to obtain the target image;

calculating a weighted average of pixels at a same position based on the sample image in the sample image group to obtain the target image; and

selecting a sample image with a highest resolution, a first frame of sample image or a last frame of sample image from the sample image group as a reference image, calculating a feature point based on the sample image in the sample image group, executing a feature point alignment operation on other sample images in the sample image group based on the reference image, and performing multi-frame image fusion on the sample images subjected to the alignment operation to obtain the target image.

7. A multi-frame image denoising method, comprising:

collecting an image group to be processed including multiple continuous frames of images through a shooting device, and acquiring a photographic sensitivity of the image group to be processed;

determining a noise characterization image corresponding to the image group to be processed based on the photographic sensitivity;

inputting the image group to be processed and the noise characterization image to a multi-frame image denoising model trained through the training method recited in claim 1 ; and

outputting a target denoised image corresponding to the image group to be processed from the multi-frame image denoising model.

8. The multi-frame image denoising method of claim 7 , wherein determining the noise characterization image corresponding to the image group to be processed based on the photographic sensitivity further comprises:

multiplying the photographic sensitivity by an adaptation coefficient input by a user to obtain an adaptively adjusted photographic sensitivity; and

determining the noise characterization image corresponding to the image group to be processed based on the adaptively adjusted photographic sensitivity.

9. The multi-frame image denoising method of claim 8 , wherein determining the noise characterization image corresponding to the image group to be processed based on the adaptively adjusted photographic sensitivity further comprises:

constructing a photographic sensitivity image where all pixel values of the photographic sensitivity image are the adaptively adjusted photographic sensitivity, and

determining the photographic sensitivity image as the noise characterization image corresponding to the image group to be processed.

10. A training device for an image denoising model, comprising:

a processor; and

a memory that is configured to store instructions executable by the processor,

wherein the processor is configured to:

collect multiple sample image groups through a shooting device, each of the sample image groups including multiple frames of sample images with a same photographic sensitivity, and sample images in different sample image groups have different photographic sensitivities;

for each of the sample image groups,

acquire a photographic sensitivity of the sample image group;

construct a photographic sensitivity image based on the photographic sensitivity and determine a noise characterization image corresponding to the sample image group based on the photographic sensitivity image, wherein all pixel values of the photographic sensitivity image are a normalized photographic sensitivity of each sample image group; and

determine a training input image group and a target image which are associated with the sample image group, the training input image group including all or part of sample images in the sample image group and the noise characterization image corresponding to the sample image group;

construct multiple training pairs, each of the training pairs including a training input image group and a corresponding target image; and

train the image denoising model based on the multiple training pairs until the image denoising model converges.

11. The training device of claim 10 , wherein the processor is further configured to, for each of the sample image groups, determine the noise characterization image corresponding to the sample image group based on the photographic sensitivity image by:

determining the photographic sensitivity image as the noise characterization image.

12. The training device of claim 10 , wherein the processor is further configured to, for each of the sample image groups, determine the noise characterization image corresponding to the sample image group based on the photographic sensitivity image by:

for each of the sample image groups, acquiring a grayscale component image of the target image corresponding to the sample image group,

determining a dot product matrix of the grayscale component image and the photographic sensitivity image, and

determining the dot product matrix as the noise characterization image.

13. The training device of claim 10 , wherein the processor is further configured to, for each of the sample image groups, determine the noise characterization image corresponding to the sample image group based on the photographic sensitivity image by:

for each of the sample image groups, determining a grayscale component image of the target image corresponding to the sample image group, and

determining that the noise characterization image is a two-dimensional image including the photographic sensitivity image and the grayscale component image.

14. The training device of claim 10 , wherein the process is further configured to calculate the photographic sensitivity by:

making statistics on a maximum photographic sensitivity and a minimum photographic sensitivity in the multiple sample image groups, determining a photographic sensitivity interval based on the maximum photographic sensitivity and the minimum photographic sensitivity, constructing a mapping relationship between the photographic sensitivity interval and an image pixel interval, and for each of the sample image groups, determining the normalized photographic sensitivity of the sample image group according to the mapping relationship.

15. The training device of claim 10 , wherein the processor is further configured to determine the target image associated with the sample image group by one of:

for each of the sample image groups,

calculating an average of pixels at a same position based on a sample image in the sample image group to obtain the target image;

calculating a weighted average of pixels at a same position based on the sample image in the sample image group to obtain the target image; and

selecting a sample image with a highest resolution, a first frame of sample image or a last frame of sample image from the sample image group as a reference image, calculating a feature point based on the sample image in the sample image group, executing a feature point alignment operation on the other sample images in the sample image group based on the reference image, and performing multi-frame image fusion on the sample images subjected to the alignment operation to obtain the target image.

16. A multi-frame image denoising device, comprising:

a processor; and

a memory configured to store instructions executable by the processor,

wherein the processor is configured to implement the method of claim 7 .

17. The multi-frame image denoising device of claim 16 , wherein the processor is further configured to determine the noise characterization image corresponding to the image group to be processed based on the photographic sensitivity by:

multiplying the photographic sensitivity by an adaptation coefficient input by a user to obtain an adaptively adjusted photographic sensitivity, and

determining the noise characterization image corresponding to the image group to be processed based on the adaptively adjusted photographic sensitivity.

18. The multi-frame image denoising device of claim 17 , wherein determining the noise characterization image corresponding to the image group to be processed based on the adaptively adjusted photographic sensitivity further comprises:

constructing a photographic sensitivity image where all pixel values of the photographic sensitivity image are the adaptively adjusted photographic sensitivity, and

determining the photographic sensitivity image as the noise characterization image corresponding to the image group to be processed.

19. A non-transitory computer-readable storage medium having stored executable instructions thereon that, when executed by a processor, implement the steps of the method of claim 1 .

20. A non-transitory computer-readable storage medium having stored executable instructions thereon that, when executed by a processor, implement the steps of the method of claim 7 .