An image denoising method based on Gram angle field and selected state space model

Abstract

The present invention provides an image denoising method based on a Gram angular field and a selected state space model, which relates to the technical field of image denoising, and includes: performing image extraction on an image to be denoised to obtain various image blocks, and determining a Gram angular field matrix according to the angles between the image blocks; determining an enhanced feature vector according to the image blocks and the corresponding Gram angular field matrix; and inputting the enhanced feature vector into a trained selected state space model to obtain a denoised image. Compared with traditional image denoising methods, the present invention can reduce redundant calculations and improve overall efficiency by calculating the Gram angular field when calculating the geometric relationship between image blocks.

Description

Technical Field

[0001] The invention relates to the technical field of image denoising, and in particular to an image denoising method based on a Gram angle field and a selected state space model. Background Art

[0002] In the field of image denoising, existing technologies are mainly divided into the following categories:

[0003] 1. Traditional filtering method:

[0004] Mean filtering and median filtering remove noise by calculating the average or median value of local pixels. These methods are simple and easy to use, but they can easily blur image details. Gaussian filtering uses a Gaussian kernel for convolution operations, which can effectively remove Gaussian noise, but has limited ability to retain details.

[0005] 2. Transform domain method:

[0006] Wavelet transform converts the image to the wavelet domain, removes noise through threshold processing, and then performs inverse transform to restore the image. This method has advantages in multi-scale analysis, but the computational complexity is high.

[0007] 3. Model-based methods:

[0008] Non-local mean (NLM) algorithm: removes noise by calculating the weighted average of non-local similar blocks in the image. The effect is good, but the computational complexity is high. BM3D (Block-Matching and 3D filtering) algorithm stacks similar image blocks into a three-dimensional array, removes noise in the transform domain, and then reconstructs the image. BM3D has excellent results, but the computational complexity is still high.

[0009] 4. Deep Learning Methods:

[0010] Convolutional Neural Network (CNN): Uses deep learning models to learn denoising features through a large amount of training data. Although the effect is significant, the training time is long and the model architecture and parameter selection are demanding. Generative Adversarial Network (GAN): Denoising is achieved through adversarial training of the generative network and the discriminative network, but the GAN model is difficult to train and has poor stability.

[0011] Technical problems in existing solutions

[0012] High computational complexity. Many advanced denoising algorithms, such as BM3D and non-local means algorithm, have high computational complexity and slow processing speed, although they have good denoising effects, and are difficult to meet the needs of real-time applications.

[0013] Insufficient detail preservation. Traditional filtering methods, such as mean filtering and Gaussian filtering, easily blur image details during the denoising process, resulting in degraded image quality.

[0014] Model training is complex. Deep learning methods rely on a large amount of training data and computing resources. The training time is long and the requirements for model architecture and parameter selection are high, so their applicability is limited.

[0015] Poor robustness: some existing methods perform inconsistently when dealing with different types of noise and lack versatility. In particular, deep learning models may rely heavily on training data, resulting in poor robustness in practical applications. Summary of the invention

[0016] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide an image denoising method based on Gram angle field and selected state space model.

[0017] To achieve the above object, the present invention provides the following solutions:

[0018] An image denoising method based on Gram angle field and selected state space model, comprising:

[0019] Performing image extraction on the image to be denoised to obtain image blocks, and determining a Gram angular field matrix according to the angles between the image blocks;

[0020] Determine an enhanced feature vector according to the image block and the corresponding Gram angular field matrix;

[0021] The enhanced feature vector is input into the trained selection state space model to obtain a denoised image.

[0022] Preferably, the image to be denoised is subjected to image extraction to obtain image blocks, and a Gram angular field matrix is ​​determined according to the angles between the image blocks, including:

[0023] The denoised image of size M×N is divided into image blocks of size k×k; the set of image blocks after division is represented as B={b1,b2,...,b T},in, Wherein, M and N are the length and width of the image to be denoised, respectively, k represents the number of pixels contained in each image block, T is the total number of image blocks, B is the set of image blocks, and b T is the T-th image block;

[0024] For each pair of image blocks b i and b j , calculate the angle θ between them ij θ ij The calculation formula is: Among them, b i b j Vector b representing the image block i and bj The inner product of i ‖ and ‖b j ‖ respectively represent vector b i and b j The norm of ;

[0025] The angle θ between all image patches ij The elements Θ in the Gram angular field matrix Θ are composed of ij Vector b representing the image block i and b j The angle between.

[0026] Preferably, determining an enhanced feature vector according to the image block and the corresponding Gram angular field matrix comprises:

[0027] Compare the original pixel value of each image block with the feature Θ of the corresponding Gram angular field matrix Θ ij Combine to form the initial eigenvector f i :The calculation formula of the initial eigenvector is: i =[b i ,Θ i1 ,Θ i2 ,...,Θ iT ]; where Θ ij Represents image block b i With other image blocks b j The angle between

[0028] The initial feature vectors are normalized and concatenated into a unified enhanced feature vector.

[0029] Preferably, the enhanced feature vector is input into a trained selection state space model to obtain a denoised image, comprising:

[0030] Determine the architecture of a selection state space model; the selection state space model includes an input layer, a plurality of selection state layers and an output layer;

[0031] The selected state space model is trained using the preset enhanced feature vector to obtain a trained selected state space model; during the training process, the parameters of the selected state space model are optimized by minimizing the error between the denoised image block and the noise-free image block;

[0032] The enhanced feature vector of the image to be denoised is input into the trained selection state space model to predict the denoised image patch through the dynamic selection mechanism of the selection state layer;

[0033] The denoised image blocks are reassembled to form a complete denoised image.

[0034] Preferably, the formula for state selection and information transmission in the state selection layer is:

[0035] h t+1 =S k A k h t +b k ; Among them, h t Indicates the current state, h t+1 Indicates the next state, A k is the state transfer matrix, S k is the state selection matrix, b k is the bias term.

[0036] Preferably, the error function of the selected state-space model is for: in, is the denoised image block, b i is a noise-free image patch.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] The present invention provides an image denoising method based on the Gram angular field and the selected state space model, comprising: performing image extraction on the image to be denoised to obtain each image block, and determining a Gram angular field matrix according to the angle between the image blocks; determining an enhanced feature vector according to the image block and the corresponding Gram angular field matrix; inputting the enhanced feature vector into the trained selected state space model to obtain a denoised image. Compared with the traditional image denoising method, the present invention can reduce redundant calculations and improve overall efficiency by calculating the Gram angular field when calculating the geometric relationship between image blocks. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0040] Figure 1 A flow chart of a method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides an image denoising method based on Gram angle field and selected state space model, comprising:

[0044] Step 100: performing image extraction on the image to be denoised to obtain image blocks, and determining a Gram angular field matrix according to the angles between the image blocks;

[0045] Step 200: determining an enhanced feature vector according to the image block and the corresponding Gram angular field matrix;

[0046] Step 300: Input the enhanced feature vector into a trained selection state space model to obtain a denoised image.

[0047] Specifically, the steps of the method for extracting geometric relationships of image blocks based on the Gram angle field in this embodiment are as follows:

[0048] Step 1.1: Image patch extraction

[0049] The input image I is divided into small blocks of size k×k. For an image of size M×N, the divided image blocks can be represented as B={b1,b2,...,b T},in

[0050] Parameter k:

[0051] Meaning: The size of the image block, that is, the number of pixels contained in each image block. Specifically, the image is divided into small blocks of size k×k.

[0052] Function: The basic unit for image segmentation. Choosing an appropriate k value can balance image details and computational complexity.

[0053] T:

[0054] It means the total number of image blocks. For an image of size M×N, T represents the total number of image blocks after segmentation.

[0055] Calculation formula:

[0056] Function: Indicates the number of small blocks into which the entire image is divided. These small blocks will be used for subsequent image denoising processing.

[0057] Step 1.2: Calculate the angle between image patches

[0058] For each pair of image blocks b i and b j , calculate the angle θ between them ij , the formula is as follows:

[0059]

[0060] Among them, b i b j Represents vector b i and b j The inner product of i ‖ and ‖b j ‖ respectively represent vector b i and b j The norm of .

[0061] Step 1.3: Form the Gram angular field matrix

[0062] The angle θ between all image patches ij The elements Θ of the matrix Θ are composed of ij Represents image block b i and b j The angle between.

[0063] Furthermore, the steps of the enhanced feature vector construction method in this embodiment combined with the Gram angular field feature are as follows:

[0064] Step 2.1: Feature vector construction

[0065] The original pixel value b of each image block i The corresponding Gram angular field feature Θ ij Combined to form a new enhanced feature vector f i :

[0066] f i =[b i ,Θ i1 ,Θ i2 ,...,Θ iT ]

[0067] Among them, Θ ij Represents image block b i With other image blocks b j The angle between.

[0068] Step 2.2: Feature normalization and concatenation

[0069] The feature vectors are standardized to ensure their consistency and applicability. The standardized feature vectors are concatenated into a unified feature vector.

[0070] Furthermore, the steps of selecting the application and optimization of the state space model in image denoising in this embodiment are as follows:

[0071] Step 3.1: Choose a state-space model architecture design

[0072] Design the architecture of the selection state space model, including the input layer, multiple selection state layers and the output layer. The selection state layer selects the appropriate state for information transmission and processing through a dynamic selection mechanism. The specific architecture design is as follows:

[0073] Input layer: receives the enhanced feature vector f i .

[0074] Select state layer: Each layer contains several states, each state has a corresponding state transfer matrix A k and the state selection matrix S k , the state selection layer performs state selection and information transmission through the following formula:

[0075] h t+1 =S k A k h t +b k

[0076] Among them, h t Indicates the current state, h t+1 Indicates the next state, A k is the state transfer matrix, S k is the state selection matrix, b k is the bias term.

[0077] Output layer: Outputs the denoised image blocks.

[0078] Step 3.2: Model training

[0079] Using the constructed enhanced feature vector {f1,f2,...,f T}Training selects the state space model. During the training process, the denoised image blocks are minimized With the noise-free image patch b i The error between is used to optimize the model parameters. The error function is as follows:

[0080]

[0081] Step 3.3: Calculate the Gram angular field features of the image to be denoised

[0082] In the actual image denoising process, the image to be denoised is divided into small blocks, and the angle between each image block is calculated to form the Gram angle field feature matrix Θ of the image to be denoised.

[0083] Step 3.4: Image denoising

[0084] The enhanced feature vector {f1,f2,...,f T} is input into the trained selection state space model, and the model predicts the denoised image block through the dynamic selection mechanism of the selection state layer

[0085] In both steps, the concatenated feature vectors are used as input data for training the model and performing denoising.

[0086] Step 3.5: Image patch stitching

[0087] The denoised image block Reassemble into a complete denoised image.

[0088] Specifically, the advantages of each step of this embodiment are as follows:

[0089] Steps 1.2 and 1.3: Ensure that the geometric relationship between image patches is accurately represented, which helps in the subsequent feature enhancement.

[0090] Steps 2.1 and 2.2: Construct an enhanced feature vector containing the original information and geometric relationship of the image block to improve the feature representation ability of the model.

[0091] Step 3.1: Design and select the architecture of the state-space model to improve the flexibility and adaptability of the model through a dynamic selection mechanism.

[0092] Step 3.2: Select the state-space model through training and learn an effective denoising strategy.

[0093] Step 3.4: Use the trained model to denoise the image to ensure the robustness and accuracy of the final denoising effect.

[0094] The beneficial effects of the present invention are as follows:

[0095] (1) High efficiency: Compared with traditional image denoising methods, the image denoising method based on the Gram angle field and the selected state space model can reduce redundant calculations and improve overall efficiency by calculating the geometric relationship between image blocks through the calculation of the Gram angle field. Advantages: Steps 1.2 and 1.3 calculate the angle between image blocks and form a Gram angle field matrix to effectively capture the geometric relationship between image blocks. This step uses vector inner product and angle calculation to reduce the high-dimensional image data into a geometric relationship representation, reducing computational complexity and improving processing efficiency.

[0096] (2) High precision: The enhanced feature vector combined with the Gram angular field feature can more comprehensively describe the features of the image block, including its geometric relationship and original pixel value information, thereby improving the accuracy of denoising. Advantage source: Steps 2.1 and 2.2 construct an enhanced feature vector that combines the original pixel value of the image block with the Gram angular field feature. By increasing the feature dimension, the model can capture richer image information, including texture and geometric relationships, thereby improving the accuracy of the denoising model.

[0097] (3) Strong robustness: The application of the state-space model in image denoising can adaptively handle different types of noise through a dynamic selection mechanism, ensuring the stability and consistency of the denoising results. Advantages: In steps 3.1 and 3.3, the state-space model is trained and selected, and the processing method that adapts to different noise characteristics is dynamically selected during the denoising process. The model learns the noise characteristics in the training data and optimizes the model parameters, so that it can adaptively handle various types of noise in practical applications, ensuring the stability and consistency of the denoising results.

[0098] In summary, the present invention improves processing efficiency by calculating the angle between image blocks and forming a Gram angular field matrix, and by combining the original pixel values ​​of the image blocks with the Gram angular field features to construct an enhanced feature vector, enriches the feature representation, and improves the denoising accuracy. By training the selected state space model and dynamically selecting a processing method that adapts to different noise features during the denoising process, the robustness and reliability of the model are enhanced. These advantages make the image denoising method based on the Gram angular field and the selected state space model have significant advantages over the prior art in terms of processing efficiency, denoising accuracy, and processing robustness.

[0099] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0100] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An image denoising method based on Gram angle field and selected state space model, characterized in that: include: Performing image extraction on the image to be denoised to obtain image blocks, and determining a Gram angular field matrix according to the angles between the image blocks; Determine an enhanced feature vector according to the image block and the Gram angular field matrix; Inputting the enhanced feature vector into a trained selection state space model to obtain a denoised image; Determining an enhanced feature vector according to the image block and the corresponding Gram angular field matrix includes: Compare the original pixel value of each image block with the feature Θ of the corresponding Gram angular field matrix Θ ij Combined to form the initial eigenvector f i :The calculation formula of the initial eigenvector is: i =[b i ,Θ i1 ,Θ i2 ,...,Θ iT ]; where Θ ij The vector representation b of the image block i Vector representation b of other image patches j The angle between them; T is the total number of image blocks; The initial feature vectors are normalized and concatenated into a unified enhanced feature vector.

2. The image denoising method based on Gram angle field and selected state space model according to claim 1, characterized in that: The image to be denoised is subjected to image extraction to obtain image blocks, and a Gram angle field matrix is ​​determined according to the angles between the image blocks, including: The denoised image of size M×N is divided into image blocks of size k×k; the set of image blocks after division is represented as B={b1,b2,...,b T },in, Wherein, M and N are the length and width of the image to be denoised, respectively, k represents the number of pixels contained in each image block, B is the image block set, and b T is the vector representation of the T-th image block; For each pair of image patches, the vector b i and b j , calculate the angle θ between them ij θ ij The calculation formula is: Among them, b i b j The vector representation b of the image block i and b j The inner product of i ‖ and ‖b j ‖ respectively represent vectors representing b i and b j The norm of ; The angle θ between the vector representations of all image patches ij The elements Θ in the Gram angular field matrix Θ are composed of ij The vector representation b of the image block i and b j The angle between.

3. The image denoising method based on Gram angle field and selected state space model according to claim 2, characterized in that: The enhanced feature vector is input into the trained selection state space model to obtain a denoised image, including: Determine the architecture of a selection state space model; the selection state space model includes an input layer, a plurality of selection state layers and an output layer; The selected state space model is trained using the preset enhanced feature vector to obtain a trained selected state space model; during the training process, the parameters of the selected state space model are optimized by minimizing the error between the denoised image block and the noise-free image block; The enhanced feature vector of the image to be denoised is input into the trained selection state space model to predict the denoised image patch through the dynamic selection mechanism of the selection state layer; The denoised image blocks are reassembled to form a complete denoised image.

4. The image denoising method based on Gram angle field and selected state space model according to claim 3 is characterized in that: The formula for state selection and information transmission in the state selection layer is: h t+1 =S k A k h t +b k ; Among them, h t Indicates the current state, h t+1 Indicates the next state, A k is the state transfer matrix, S k is the state selection matrix, b k is the bias term.

5. The image denoising method based on Gram angle field and selected state space model according to claim 4, characterized in that: The error function L of the selected state space model is: in, is the vector representation of the denoised image block, b i is the vector representation of the noise-free image patch.

Images