CN118096567A

CN118096567A - DBN model self-adaptive image denoising method and system based on Bayesian method

Info

Publication number: CN118096567A
Application number: CN202410215282.0A
Authority: CN
Inventors: 沈卉卉; 钱坤; 张耀峰; 黄振东; 徐勇; 杨明普
Original assignee: HUBEI UNIVERSITY OF ECONOMICS
Current assignee: HUBEI UNIVERSITY OF ECONOMICS
Priority date: 2024-02-27
Filing date: 2024-02-27
Publication date: 2024-05-28
Anticipated expiration: 2044-02-27
Also published as: CN118096567B

Abstract

The application discloses a DBN model self-adaptive image denoising method and system based on a Bayesian method, which are used for preprocessing an input noise image, including image normalization, image segmentation and the like, then establishing a priori probability and a likelihood function of an image denoising model by using the Bayesian method, and estimating a denoised image by maximizing posterior probability; then, the random noise removal is carried out on the image data by utilizing a correction weight attenuation momentum method of combining a restricted Boltzmann machine unsupervised learning model with random variable dB leaf methods in different forms, and the image denoising effect can be effectively improved by combining a Bayesian method and image denoising of a depth belief network model algorithm formed by a plurality of RBMs, so that the method has higher robustness and adaptability. Compared with the traditional image denoising method, the method has the advantages that the detail information of the image can be better reserved, denoising parameters can be flexibly adjusted according to the requirements of actual application scenes, and the optimal denoising effect is achieved.

Description

DBN model self-adaptive image denoising method and system based on Bayesian method

Technical Field

The application relates to the technical field of DBN model self-adaptive image denoising based on a Bayesian method, in particular to a DBN model self-adaptive image denoising method and system based on the Bayesian method.

Background

With the development of digital image processing technology, the image denoising technology has important significance in the fields of computer vision, medical imaging and the like. The current image denoising method based on deep learning has great advantages in performance, wherein a convolutional neural network (Convolutional Neural Networks, CNN), a generating countermeasure network (GENERATIVE ADVERSARIAL network, GAN) and a traditional unsupervised learning algorithm have the best denoising effect, and a block matching (BM3D) algorithm has excellent performance in the field of image denoising. However, the CNN model denoising method and the traditional unsupervised learning BM3D algorithm both need a large amount of training sets of noiseless image data for training, and the lack of the training sets and the like can affect the denoising effect and are difficult to popularize in actual data application, and the training is often carried out by the GPU due to the scale of training samples or the need of denoising the CNN model, which takes a very long time. In addition, the existing limited Boltzmann machine (RESTRICTED BOLTZMANN MACHINE, RBM) model image denoising method is completely based on an RBM undirected graph model, the distribution of a lower layer also depends on the distribution of a higher layer, the calculation is complex, the denoising effect is also general, and the method has poor adaptability to noise types and intensity changes.

Image denoising is an important task in image processing, and has important roles in many application fields, such as computer vision, medical imaging, satellite remote sensing, seismic exploration, and the like. However, the conventional image denoising method often has the problems of unsatisfactory denoising effect, sensitivity to noise and the like, and the deep learning denoising algorithm often needs a large amount of training data to learn the priori knowledge of the image, which is sometimes difficult to meet in practical application. Second, these methods lack adaptivity to noise type and intensity, possibly resulting in poor denoising.

Disclosure of Invention

The application provides a DBN model self-adaptive image denoising method and system based on a Bayesian method, and aims to solve the problems of insufficient training data, long training time, high computational complexity, general denoising effect, poor adaptability to different types and intensities of noise and the like in the prior art.

In a first aspect, a method for denoising a DBN model adaptive image based on a bayesian method, the method comprising:

Training the RBM model by using a Bayesian method, and introducing priori knowledge and uncertainty information in the training to obtain a trained RBM model;

Constructing DBN models of 3 hidden layers by using the trained RBM, and denoising the target noisy image;

The denoising process includes: preprocessing the target noisy image, and extracting the image characteristics; the pretreatment comprises the following steps: normalizing and image segmentation processing is carried out on the target noisy image;

According to the image characteristics, establishing prior probability and likelihood function of the DBN model of image denoising by using a Bayesian method;

Analyzing the noise type and intensity according to the noise level of the image, and adjusting the Bayesian RBM model parameters;

Denoising the preprocessed image by using the adjusted Bayesian DBN model to obtain a denoised image;

And carrying out post-processing on the denoised image to obtain a target denoised image.

In the above solution, optionally, the training the RBM model by using a bayesian method, and introducing priori knowledge and uncertainty information into the training to obtain each trained RBM model includes:

The method is realized through a Bayesian inference algorithm, and knowledge of domain experts or previous research results are utilized to integrate the priori knowledge into a model training process; and incorporates uncertainty in model parameters into the model training and inference process.

In the above solution, optionally, the integrating the prior knowledge into the model training process by using knowledge of domain experts or previous research results specifically includes:

the prior knowledge includes knowledge of image feature distribution, noise type, and intensity; the priori knowledge is used for defining the prior probability distribution of the model or restricting the value range of the model parameters;

In the RBM model, the structures, connection modes and the value ranges of weights of the visible layer and the hidden layer are set through priori knowledge.

In the above solution, optionally, the step of incorporating uncertainty of model parameters into model training and inference process specifically includes: uncertainty in the parameters is described by introducing posterior probability distributions using a variational bayesian approach.

In the above solution, optionally, preprocessing the target noisy image includes: and carrying out normalization and image segmentation processing on the target noisy image.

In a second aspect, a system for denoising a DBN model adaptive image based on a bayesian method, the system comprising:

Training module: the method is used for training the RBM model by using a Bayesian method, and introducing priori knowledge and uncertainty information in the training to obtain a trained RBM model;

And a denoising module: the method comprises the steps of constructing DBN models of 3 hidden layers by using a trained RBM, and denoising a target noisy image;

denoising the preprocessed image by using the adjusted Bayesian RBM model to obtain a denoised image;

In the above solution, optionally, preprocessing the target noisy image includes: and carrying out normalization and image segmentation processing on the target noisy image. Compared with the prior art, the application has at least the following beneficial effects:

Based on further analysis and research on the problems of the prior art, the application realizes that the problems of insufficient training data, long training time, high calculation complexity, general denoising effect, poor adaptability to different types and intensities of noise and the like exist in the prior art. The Bayesian method can effectively process uncertain factors in the image denoising process by establishing a probability model. The RBM is used as a deep learning model, and can realize automatic noise filtering through learning the local features of the image. The Bayesian method is combined with DBN model algorithms of 3 RBM hidden layers, so that the image denoising effect can be effectively improved, and the method has higher robustness and adaptability. The method comprises the following steps: firstly, preprocessing an input noise image, including image normalization, image segmentation and the like; then, establishing a priori probability and a likelihood function of an image denoising model by using a Bayes method, and estimating a denoised image by maximizing posterior probability; then, carrying out random noise removal on the image data by using a correction weight attenuation momentum method of combining RBM (radial basis function) unsupervised learning models with random variable dB leaf methods in different forms so as to improve the denoising effect; finally, the processed image is applied to practical application scenes, such as medical images, satellite images, seismic data processing and the like.

The RBM model based on the Bayesian method is utilized to carry out self-adaptive image denoising, a large amount of noiseless image data is not needed, and the noiseless image is estimated through an iterative optimization process. Meanwhile, according to the scheme, the noise type and the intensity are analyzed according to the noise level of the image, and the parameters of each Bayesian RBM model are adaptively adjusted, so that the noise of different types and intensities can be removed. Finally, the scheme obtains the final denoised image by post-processing the denoised image, improves the quality and efficiency of image denoising, and better meets the actual application requirements.

Drawings

FIG. 1 is a schematic flow chart of a Bayesian method-based DBN model adaptive image denoising method according to an embodiment of the present application;

FIG. 2 is a flowchart of a DBN image denoising method based on a Bayesian method according to an embodiment of the present application;

fig. 3 is a change situation of PSNR values after denoising real seismic data by using a DBN model adaptive image denoising method based on a bayesian method according to an embodiment of the present application.

Detailed Description

The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.

The application provides a DBN model self-adaptive image denoising method based on a Bayesian method, which comprises the following steps:

In this embodiment, the training the RBM model using the bayesian method, and introducing priori knowledge and uncertainty information into the training to obtain a trained RBM model includes:

In this embodiment, the prior knowledge is integrated into the model training process by using knowledge of domain experts or previous research results, specifically:

In this embodiment, the step of incorporating uncertainty of model parameters into model training and inference process is specifically as follows: uncertainty in the parameters is described by introducing posterior probability distributions using a variational bayesian approach.

In this embodiment, preprocessing the target noisy image includes: and carrying out normalization and image segmentation processing on the target noisy image.

The RBM can learn probability distribution of image data unsupervised, and a Deep Belief Network (DBN) with a plurality of RBMs overlapped is a fast and effective Deep learning method. The method does not need a large number of labels, has low requirement on computer hardware, and can achieve better recognition effect in a short time. And when the DBN model is used for image denoising, the method has the advantages that a training set is not needed, pre-training is not needed, and the denoising is efficient and good in effect.

The bayesian method is a statistical inference method of a system in statistics, and is also one of core methods of machine learning, and it regards all unknown parameters as random variables obeying a certain probability distribution, and then calculates and derives a posterior distribution of the unknown parameters according to sample information provided by observation data and prior knowledge of the unknown parameters. The bayesian method has many applications in the field of machine learning, from supervised learning to unsupervised and semi-supervised learning, etc., and is almost used for any kind of learning task. In an unsupervised learning RBM model, a variation reasoning method in an RBM algorithm, a Monte Carlo method in RBM model learning, a Gibbs sampling and regularization method and a Bayesian statistical analysis method are all used.

The embodiment provides a DBN model self-adaptive image denoising method based on a Bayesian method, which aims to solve the problems of poor noise removal effect, poor adaptability and the like in the existing image denoising technology. By the method, the quality and the efficiency of image denoising can be improved, and the actual application requirements can be better met.

In one embodiment, the following technical scheme is adopted to solve the defects existing in the existing scheme:

DBN model based on Bayesian method: training the DBN models of the 3 hidden layers by using a Bayesian method, introducing priori knowledge and uncertainty information, and improving generalization capability of the model and adaptability to noise;

The RBM model is utilized to carry out self-adaptive image denoising, a large amount of noiseless image data is not needed, denoising processing can be carried out on the noiseless image, and the noiseless image is estimated through an iterative optimization process;

self-adaptive image denoising: the parameters of the RBM model are adaptively adjusted by analyzing the local characteristics and noise characteristics of the image, so that the noise of different types and intensities can be removed.

The method comprises the following specific steps:

preprocessing the noisy image, including image normalization, image segmentation and the like, and extracting essential characteristics of the image.

According to the image characteristics, establishing the prior probability and likelihood function of the DBN model of image denoising by using a Bayesian method, wherein the prior probability and likelihood function comprise determining model parameters and training the model, and estimating the denoised image by maximizing posterior probability.

And analyzing the noise type and the intensity according to the noise level and the like of the image, and adaptively adjusting the parameters of the Bayesian RBM model.

And denoising the preprocessed image by using the adjusted Bayesian DBN model.

Post-processing is carried out on the denoised image, including image reconstruction and the like, so as to obtain a final denoised image.

The denoising core algorithm of this embodiment has the following process:

Input: adding random noise with the standard deviation sigma into an original image y with the size of m multiplied by r, taking the image as x, dividing the image into l multiplied by l small block images x ', pulling the small block images into column vectors with the dimension of l ², inputting the column vectors into RBM in batches, wherein each batch has T small images, the RBM learning rate is eta, the momentum is m ^*, and the learning rate is eta ' and the momentum is m ' when the whole network is finely tuned;

Constructing 3 hidden layer prior information DBN network structures, namely, l ²-400-400-400-l², and extracting the characteristics of input x';

the RBM model algorithm of the modified weight attenuation momentum method of the random variable decibel leaf method in different forms is adopted, and an original image y is used as a label;

after the small image x 'passes through an RBM algorithm, a learned characteristic diagram f (x') is obtained;

performing splicing operation on each learned feature image, gradually recovering the feature images to the feature images with the same size as the original image, and enabling the feature images f (x) and the labels y which are spliced last to be as small as possible during fine adjustment;

And outputting to obtain the denoised image f (x) with optimal denoised performance and m multiplied by r.

Compared with the prior art, the embodiment has the following advantages and positive effects:

The training time is reduced: the DBN unsupervised learning models of the 3 hidden layers are used for denoising without training sets or training; optimizing the RBM model by a Bayesian method, and reducing the requirement for a large amount of noise image data, thereby reducing the training time; an improved momentum algorithm is adopted to improve the network denoising efficiency. By combining the self characteristics of gradient rising and gradient falling algorithms during RBM training and fine tuning, the correction weight attenuation motion items of random variable decibel leaf methods in different forms are adopted in sections to accelerate network convergence and accurately find the optimal point.

And the denoising effect is improved: by introducing a Bayesian method and priori knowledge, the invention can better capture the inherent structure and noise characteristics of the image and realize more effective noise removal. The Bayesian method can effectively adjust model parameters, combines priori knowledge and noise statistics characteristics, can effectively process uncertainty in the image denoising process, and improves generalization capability of the model, thereby preventing the occurrence of the over-fitting phenomenon and improving denoising quality.

Self-adaptive denoising: the RBM model is utilized to learn the local characteristics of the image, the self-adaptability of the Bayesian method is combined, and a large amount of training of noiseless image data is not needed, so that the calculation complexity is reduced, and the accurate filtering of noise is realized. According to the noise type and intensity, the parameters of the Bayesian RBM model are adaptively adjusted, so that the Bayesian RBM model has better adaptability and generalization capability, wider applicable scenes and higher denoising effect.

According to the image denoising method and device, the image denoising effect can be effectively improved by combining a Bayesian method and an RBM model algorithm, and the robustness and the adaptability are high. Compared with the traditional image denoising method, the embodiment can better keep the detail information of the image and has stronger noise suppression capability. In addition, the embodiment can flexibly adjust the denoising parameters according to the requirements of actual application scenes so as to achieve the optimal denoising effect.

In one embodiment, assuming a real seismic data image contaminated with noise is input, the noise image is first preprocessed, including image normalization, image segmentation, and the like. Then, a Bayes method is utilized to establish prior probability and likelihood function of the image denoising model, and the denoised image is estimated by maximizing posterior probability. And then, further processing the denoised image by using a DBN image denoising algorithm of which the RBM forms 3 hidden layers so as to improve the denoising effect. Finally, the processed image is used for seismic exploration.

As shown in fig. 2-3, the images with noise are segmented and input into RBM models in batches, and the problem of RBM network learning efficiency is solved by combining the correction weight attenuation dynamic terms of different random variable decibel leaf methods, so that RBM denoising efficiency is improved, and the method is applied to seismic data denoising of seismic exploration.

Fig. 2 illustrates a flow of a DBN image denoising method based on a bayesian method. Firstly, respectively inputting random noise images x with noise standard deviation sigma= 15,25,50; then, dividing the image into l multiplied by l small block images x 'through a preprocessing step, setting the number of visible units as l ² and the number of hidden units of each layer as n, and inputting the divided small block images x' into an RBM model in batches; then, training the local features of each RBM model learning image by using a Bayesian method, identifying according to the noise type and the intensity, adaptively adjusting the parameters of each RBM model, and carrying out denoising treatment; finally, the denoised image f (x) is output.

Under the condition that random noise sigma=25 is generated in the real seismic data, the PSNR value of the restoration map after noise is removed by using a DBN model self-adaptive image denoising method based on a Bayesian method is changed along with the iteration number, as shown in figure 3,

Through the brief description of the drawings, the embodiment provides a DBN model self-adaptive image denoising method based on a Bayesian method, which has a good denoising effect, and can flexibly and adaptively adjust denoising parameters according to actual application scene requirements so as to achieve the optimal denoising effect.

In one embodiment, a Bayesian method-based DBN model adaptive image denoising system is provided, wherein:

and a denoising module: the method comprises the steps of forming DBN models of 3 hidden layers by using trained RBMs, and denoising a target noisy image;

In this embodiment, the training the RBM model using the bayesian method, and introducing priori knowledge and uncertainty information into the training to obtain each trained RBM model includes:

In this embodiment, the step of integrating the priori knowledge into the model training process by using knowledge of domain experts or previous research results is specifically:

In this embodiment, the step of incorporating uncertainty of model parameters into model training and inference processes is specifically as follows: uncertainty in the parameters is described by introducing posterior probability distributions using a variational bayesian approach.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

Claims

1. A method for denoising a DBN model adaptive image based on a bayesian method, the method comprising:

Training a RBM model of a limited Boltzmann machine by using a Bayesian method, and introducing priori knowledge and uncertainty information into the training to obtain a trained RBM model;

constructing a depth belief network DBN model of 3 hidden layers by using the trained RBM model, and denoising the target noisy image;

2. The method of claim 1, wherein the training the RBM models using bayesian methods and introducing prior knowledge and uncertainty information into the training results in each RBM model being trained, comprising:

3. The method according to claim 2, wherein the integrating the prior knowledge into the model training process using the knowledge of the domain expert or previous research effort is specifically:

in the DBN model, the structures, the connection modes and the value ranges of the weights of the visible layer and the hidden layer are set through priori knowledge.

4. The method according to claim 2, wherein said incorporating the uncertainty of the model parameters into the model training and inference process is in particular: uncertainty in the parameters is described by introducing posterior probability distributions using a variational bayesian approach.

5. The method of claim 1, wherein preprocessing the target noisy image comprises: and carrying out normalization and image segmentation processing on the target noisy image.

6. A DBN model adaptive image denoising system based on a bayesian method, the system comprising:

7. The system of claim 6 wherein the training the RBM models using bayesian methods and introducing prior knowledge and uncertainty information into the training results in each RBM model being trained, comprising:

8. The system according to claim 7, wherein the integrating the prior knowledge into the model training process using the knowledge of the domain expert or previous research effort is specifically:

In each RBM model, the structures, connection modes and the weight value ranges of the visible layer and the hidden layer are set through priori knowledge.

9. The system according to claim 7, wherein the incorporating of the uncertainty of the model parameters into the model training and inference process is specifically: uncertainty in the parameters is described by introducing posterior probability distributions using a variational bayesian approach.

10. The system of claim 6, wherein preprocessing the target noisy image comprises: and carrying out normalization and image segmentation processing on the target noisy image.