CN-122023177-A - Image denoising method and system based on physical information guidance

CN122023177ACN 122023177 ACN122023177 ACN 122023177ACN-122023177-A

Abstract

The invention belongs to the technical field of image denoising, and aims to solve the problems of weak generalization capability, limited denoising and the like of the traditional fundus image speckle noise suppression method, and provides an image denoising method and system based on physical information guidance, wherein a registered average living fundus image and a pseudo-eye static average image are randomly fused to form a feature map to train a speckle noise estimation network; in the training process, the estimation loss is calculated based on the prediction result of the speckle noise estimation network and the mask region characteristics of the false eye still average image with the stable signal-to-noise ratio, and the noise distribution perception loss is calculated based on the feature images respectively corresponding to the prediction result of the speckle noise estimation network and the false eye still average image with the stable signal-to-noise ratio, so that the training of the speckle noise estimation network is guided. The invention uses the high signal-to-noise ratio pseudoeye still image as physical information to guide the speckle noise estimation network to realize the high signal-to-noise ratio denoising under different noise distributions.

Inventors

SONG WEIYE
WANG YUKUAN
ZHOU LIBO
CUI YUAN
QI MIN
WU FUWANG
LU XIAOQI
SHENG PENG

Assignees

山东大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (16)

1. The image denoising method based on physical information guidance is characterized by comprising the following steps: inputting the living fundus image into a trained speckle noise estimation network to obtain a speckle noise prediction result of the living fundus image; wherein, construct the data set image on the basis of the living body fundus image under different noise modes, incorporate the artificial eye still average image under different noise modes into the data set image; In the training process, the estimated loss is calculated based on the prediction result of the speckle noise estimation network and the mask region characteristics of the false eye still average image with the stable signal-to-noise ratio, and the noise distribution perceived loss is calculated based on the prediction result of the speckle noise estimation network and the characteristic images corresponding to the false eye still average image with the stable signal-to-noise ratio respectively, so that the training of the speckle noise estimation network is guided.
2. The physical information guidance-based image denoising method according to claim 1, wherein a dataset image is constructed based on live fundus images in different noise modes, and an average image of a pseudoeye still image is incorporated into the dataset image, specifically: respectively acquiring living fundus acquired images under different total frames; performing image registration on the living fundus images under different total frames, overlapping the living fundus images subjected to the image registration along a channel, and taking an average value as a dataset image; An artificial eye still image at the same total frame number as the living fundus acquired image is acquired, and an average image of the artificial eye still image is incorporated into the dataset image.
3. The method for denoising an image based on physical information guidance according to claim 1, wherein the processing procedure of the speckle noise estimation network on the input image is: Expanding the channel dimension of the input image by using a first convolution and performing feature mapping; weighting the surrounding pixel contributions of the feature map after the first convolution processing through hole convolution; And carrying out feature extraction on the feature map after the cavity convolution processing by utilizing the second convolution, and then carrying out channel compression on the feature map subjected to the feature extraction to obtain a speckle noise prediction result of the input image.
4. The image denoising method based on physical information guidance according to claim 1, wherein the living fundus image or the pseudoeye still image under different noise modes is acquired, specifically, the living fundus image or the pseudoeye still image under the corresponding noise modes under different total frame numbers is acquired.
5. The method for denoising images based on physical information guidance according to claim 1, wherein the registration average living fundus image and the artificial eye still average image are randomly fused into a Zhang Tezheng image, the fused feature image is input into the speckle noise estimation network for training, in the training process, the estimated loss is calculated by selecting the features of the 1-Mask region from the predicted value of the speckle noise estimation network and the artificial eye still average image with the signal to noise ratio tending to be stable, and the training of the speckle noise estimation network is guided by the estimated loss.
6. The method for denoising an image based on physical information guidance according to claim 1, wherein the prediction result based on the speckle noise estimation network and the pseudo-eye still image are trained using a learning method of a non-tag pair.
7. The image denoising method based on physical information guidance according to claim 1 or 6, wherein the registration average living fundus image and the artificial eye still average image are randomly fused, the fusion feature image is input into the speckle noise estimation network for training, in the training process, sampling feature images under different multiplying powers are respectively extracted from the predicted value of the speckle noise estimation network and the artificial eye still average image with stable signal-to-noise ratio, the noise distribution perception loss is calculated in one-to-one correspondence with the sampling feature images with the same scale, and the training of the speckle noise estimation network is guided through the noise distribution perception loss.
8. The method for denoising an image based on physical information guidance according to claim 1 or 5, wherein the estimated loss comprises an image gradient loss, a structural similarity loss and a separation loss, wherein the separation loss is used for realizing mapping relation and semantic information perception at a pixel level and separating characteristics of structure and noise.
9. The method for denoising an image based on physical information guidance according to claim 8, wherein the image gradient loss The method comprises the following steps: Wherein, the Is the image at the first A value of a directional gradient in the width direction at each pixel, Is the image at the first The directional gradient value in the height direction at each pixel, And Respectively representing a width direction gradient value and a height direction gradient value of a model predicted value of the fusion characteristic image, And The width direction gradient value and the height direction gradient value of the pseudo-eye still average image, which tend to stabilize the signal-to-noise ratio, are respectively represented.
10. The method for denoising an image based on physical information guidance according to claim 8, wherein the structural similarity loss is The method comprises the following steps: Wherein, the And Respectively represents the average value of the model predicted value and the average value of the false eye still average image label value with stable signal to noise ratio, And The variance of the model predicted value and the variance of the false eye still average image label value where the signal-to-noise ratio tends to be stable are respectively represented, Covariance of model predicted value and covariance of false eye still average image label value with stable signal-to-noise ratio.
11. The method for denoising an image based on physical information guidance according to claim 8, wherein the separation The loss is specifically as follows: Wherein, the A model predictor representing the fused feature image, A false eye still average image label value indicating that the signal-to-noise ratio tends to be stable, Represent the first The first batch of Line 1 The predicted value of a column of pixels, Represent the first The first batch of Line 1 The column pixel tag value is used to determine, In the case of a batch size of the product, And Representing the height and width of the image, Represent the first The number of batches of the product is one, Represent the first The number of pixels in a row is, Represent the first And column pixels.
12. The method for denoising an image based on physical information guidance according to claim 1, wherein the noise distribution perceives a loss The method comprises the following steps: Wherein, the And Respectively representing the mean value of the ith channel of the model predictive feature diagram and the mean value of the ith channel of the false eye static average image with stable signal to noise ratio, And Respectively representing the standard deviation of the ith channel of the model predictive feature map and the standard deviation of the ith channel of the false eye still average image with stable signal to noise ratio, Is the total number of channels.
13. An image denoising system based on physical information guidance, comprising: the noise estimation module is configured to input the living fundus image into a trained speckle noise estimation network to obtain a speckle noise prediction result of the living fundus image; a training module configured to construct a dataset image based on the live fundus images in the different noise modes, incorporating the pseudoeye still average image in the different noise modes into the dataset image; In the training process, the estimated loss is calculated based on the prediction result of the speckle noise estimation network and the mask region characteristics of the false eye still average image with the stable signal-to-noise ratio, and the noise distribution perceived loss is calculated based on the prediction result of the speckle noise estimation network and the characteristic images corresponding to the false eye still average image with the stable signal-to-noise ratio respectively, so that the training of the speckle noise estimation network is guided.
14. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the method of any one of claims 1-12.
15. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the method of any of claims 1-12.
16. A computer program product comprising a computer program which, when executed by a processor, implements the method of any of claims 1-12.

Description

Image denoising method and system based on physical information guidance Technical Field The invention belongs to the technical field of image denoising, and particularly relates to an image denoising method and system based on physical information guidance. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. Optical Coherence Tomography (OCT) is a technique that uses weak interference signals to make three-dimensional measurements of a sample. In recent years, OCT systems have been designed to blend visible light with near infrared light VNOCT in order to achieve high resolution and large imaging depths. Currently VNOCT is a powerful tool for early diagnosis of high blindness fundus diseases. However, in VNOCT systems there is a difference in the intensity and distribution of speckle noise due to the difference in the incident eye-safe power and the phase modulation effect of the optical path device for different wavelengths. The existence of speckle noise makes the layer structure in the fundus image difficult to observe, and the layer segmentation is not accurate enough, so that clinical diagnosis is affected. At present, speckle noise suppression methods of OCT retinal images are largely classified into low-rank estimation, supervised, unsupervised, and unlabeled peer-to-peer methods. The low-rank estimation method needs to introduce an invariance assumption of an adjacent frame structure to finish registration of a plurality of B-Scans and minimize low-rank error iterative estimation to realize denoising, and has weak generalization capability and long calculation time; the monitoring method is to obtain a plurality of B-Scan average high-definition images as true values for training of label pairs, the upper limit of performance is limited by an image registration average method, and the non-monitoring method is to form soft label pairs through the modes of internal mining of statistical information of noise images, gaussian noise simulation and the like so as to conduct denoising, and the denoising performance is close to that of the monitoring method. The non-tag pair denoising method maps the noise distribution of the strong noise image to the high-quality image, and forms the tag pair with the high-quality image into the tag pair to supervise and denoise, so that the generalization capability is weak. In summary, the existing retina image speckle noise suppression method generally has the problems of limited generalization capability, high calculation cost, limited denoising performance by registration or label generation modes and the like. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides an image denoising method and system based on physical information guidance, which utilize a high signal-to-noise ratio pseudoeye still image as a physical information guidance speckle noise estimation network to realize high signal-to-noise ratio denoising under different noise distributions. In order to achieve the above purpose, the present invention adopts the following technical scheme: in a first aspect, the present invention provides an image denoising method based on physical information guidance, including: inputting the living fundus image into a trained speckle noise estimation network to obtain a speckle noise prediction result of the living fundus image; wherein, construct the data set image on the basis of the living body fundus image under different noise modes, incorporate the artificial eye still average image under different noise modes into the data set image; In the training process, the estimated loss is calculated based on the prediction result of the speckle noise estimation network and the mask region characteristics of the false eye still average image with the stable signal-to-noise ratio, and the noise distribution perceived loss is calculated based on the prediction result of the speckle noise estimation network and the characteristic images corresponding to the false eye still average image with the stable signal-to-noise ratio respectively, so that the training of the speckle noise estimation network is guided. In a second aspect, the present invention provides an image denoising system based on physical information guidance, comprising: the noise estimation module is configured to input the living fundus image into a trained speckle noise estimation network to obtain a speckle noise prediction result of the living fundus image; a training module configured to construct a dataset image based on the live fundus images in the different noise modes, incorporating the pseudoeye still average image in the different noise modes into the dataset image; In the training process, the estimated loss is calculated based on the prediction result of the speckle noise estimation network and the mask region characteristics of the false eye stil