A degraded image restoration method based on intelligent page turning and accelerated TV-PDE

Through the intelligent page turning machine and support vector machine classifier combined with the overall variation model, the weighting factor and anisotropic diffusion technology are used to improve the TV-PDE model, solving the difficulty of removing stains and blurs in image repair of degraded documents and the slow repair speed. The problem of this achieves fast and efficient image repair effect.

CN114926372BActive Publication Date: 2024-12-03NANJING TECH UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202210752519.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-28
Publication Date
2024-12-03
Estimated Expiration
2042-06-28

AI Technical Summary

Technical Problem

The prior art is difficult to effectively remove stains and blurs in image repair of degraded documents, and the repair speed based on partial differential equations is slow, so texture information cannot be effectively repaired.

Method used

The intelligent page turning machine is used to combine the support vector machine classifier and the overall variation model, and the TV-PDE model is improved through weighting factors and anisotropic diffusion-based noise reduction method to achieve rapid image repair.

Benefits of technology

It improves the repair speed and quality of degraded document images, can effectively repair texture information, significantly improve visual effects, and overcomes the problems of slow iteration operations and loss of texture information in the original technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114926372B_ABST
    Figure CN114926372B_ABST
Patent Text Reader

Abstract

The present invention discloses a degraded image restoration method based on intelligent page turning and accelerated TV-PDE. It includes an automatic book turning machine automatic image input module, a hybrid support vector machine classifier module of Relief-MRMR-SVM, an overall variational smoothing denoising module, an overall variational fast image restoration module and a weighted factor overall variational restoration module. The support vector machine classifier module classifies the input data set image into a stained image and a long-term degraded image, and uses overall variational smoothing to denoise the divided image. Finally, a conclusion is drawn by comparing the performance of two different types of images using improved overall variational fast images and weighted factor overall variational restoration. The present invention has a simple and intelligent structure, can greatly improve the visual effect, eliminate the gradient effect, is superior to the original TV model in the restoration of texture information, and effectively shortens the running time of the original TV model algorithm, and has a wide range of application value in image restoration and other fields.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention discloses a degraded image restoration method based on intelligent page turning and accelerated TV-PDE, and relates to the technical field of image processing, mechanical structure and machine learning. Background Art

[0002] Before the rise of computers, archives, historical documents, books, etc. were all recorded and circulated in paper form. With the rise of computer technology and the advent of the Internet age, and the improvement of living standards, digital archives are booming as an effective storage and retrieval method. Compared with traditional paper documents, electronic media are more environmentally friendly and easy to carry.

[0003] Although electronic books are already quite popular, traditional paper books still exist in large quantities. For elderly people with limited mobility and some special groups with diseases, it is difficult to turn pages and read. In order to help these special groups read, it is necessary to design a book-turning robot that can automatically turn pages, and control the page turning and the direction of the page turning through voice, touch screen, buttons and gestures to assist reading. For some precious historical documents, it is more conducive to the circulation and protection of documents. In the field of document image processing, there has been a wave of methods based on mathematics and deep learning, which are applied to the problem of automatic and accurate text binarization. The method based on partial differential equations is known for its relatively low complexity and high efficiency, while the deep learning method has greater computational and structural complexity, but it is the most accurate method. The research significance of the present invention lies mainly in how to reasonably and accurately process degraded document images through a combination of mechanical structure and machine learning, so that they can be better digitally preserved.

[0004] With the development and progress of modern society, people are paying more and more attention to the quality of life. However, as more and more technological products are added to people's lives, people are gradually moving away from traditional paper book reading, which has led to a series of problems such as myopia among teenagers and dependence on technological products. According to survey data, the most important factor for teenagers' unwillingness to accept traditional paper reading is that traditional reading cannot satisfy the interests of this group and paper reading is boring. At present, many scholars at home and abroad have also conducted research on automatic book turning devices, such as the pedal-operated book turning mechanism studied by Dalian Jiaotong University. All kinds of book turning machines have certain characteristics and can complete basic book turning actions well. They can be used in coffee shops, libraries and other occasions to increase the attractiveness of paper reading to teenagers.

[0005] Since the 1980s, the research on image processing methods based on partial differential equations has attracted the research interests of a large number of mathematicians and image scholars at home and abroad. It has accumulated rich research results and shown strong vitality. At present, image processing technology using partial differential equations has been widely used in image restoration, image segmentation, image reconstruction, image recognition, image analysis, etc. The partial differential equation model regards image information as a continuous two-dimensional function and the discrete image transformation as a continuous differential operator, which makes local nonlinear analysis easy to implement. In addition, partial differential equations have a good analytical theoretical foundation. There is already a relatively complete system in theoretical analysis and numerical calculation, which provides a guarantee for theoretical proof and experiment of image processing problems.

[0006] Since image degradation is inevitable during the process of image formation, copying, and transmission, but clear and high-quality images are required in many application fields, the degraded images must be restored and enhanced. Image restoration is to reconstruct the original image from the observed degraded image. It is the basis of image processing and pattern recognition, and has a wide range of applications in astronomy, remote sensing imaging, medical imaging, etc. At present, there are still many problems that need to be solved for degraded document image enhancement, mainly the following problems:

[0007] 1. Many target images exhibit degradations that are in the same information band as the text. Therefore, attempts to remove them can result in the elimination of important text information in the processed document image.

[0008] 2. Document image binarization preprocessing: The partial differential equation-based binarization method can only enhance or suppress the available image information, so how to choose the appropriate preprocessing steps to suppress or eliminate these degradations while improving the required text information without adverse effects becomes a difficulty.

[0009] 3. Although most partial differential equation (PDE) based models perform well overall, their performance is still hampered by degraded document images with blur, oil and ink stains, smudges, uneven lighting or shadows, etc. Summary of the invention

[0010] Image restoration technology based on partial differential equations (PDEs) basically involves the discretization and iterative solution of partial differential equations in applications. The process is abstract, complex, and slow. The key idea of ​​the basic TV image restoration model is that the problem of finding the extreme value of the functional under constraints has morphological invariance and is suitable for repairing small areas such as scratches. However, the large number of iterative operations in the restoration process will inevitably lead to shortcomings such as slow restoration speed and inability to repair texture information.

[0011] In order to realize a fully automatic degraded document image restoration system model, the present invention uses an adjustable automatic book turning machine to automatically turn pages and change books of degraded document images that need to be restored. At the same time, when the photo entry software determines the picture to be restored through the human-computer interaction interface, the image processing computer drives the camera to capture the picture, and automatically enters the photographed picture to be restored into the computer for display. A large amount of effective information of the picture to be restored after entering the computer is extracted, and a data classification scheme based on support vector machine (SVM) is proposed. The required data is divided into two categories, pictures with stains and blurred pictures. Visual and mathematical methods are used to quantify and identify features, and preliminary analysis is performed on the images of stained and blurred degraded documents. Based on the analysis of the image restoration technology theory based on the total variation TV (Total Variation, TV) model, a fast image restoration method based on the improved TV model is proposed. This method draws on the core idea of ​​the TV restoration model, and realizes the restoration of defect points in one operation by adding the restoration sequence, increasing the domain range, and improving the weighting factor and other key technologies.

[0012] In order to achieve the above object, the present invention adopts the following technical solution:

[0013] A degraded image restoration method based on intelligent page turning and accelerated TV-PDE, the specific steps are as follows:

[0014] (1) Place the image book to be restored on the intelligent book turning machine, and use the intelligent book turning machine to automatically turn pages and change books of the degraded document image book to be restored; at the same time, when the photo input software determines the image to be restored through the human-computer interaction interface, the image processing computer drives the camera to capture the image, and automatically inputs the taken image to be restored into the computer for display, thereby realizing automatic image acquisition processing;

[0015] (2) Extract effective information from the images to be restored after they are entered into the computer, use visual and mathematical methods to quantify and identify the features of stained and blurred images, and analyze and restore stained, blurred and degraded document images;

[0016] (3) The analysis and restoration method specifically includes a support vector machine classifier module, a total variation (TV) smoothing denoising module, a weighted factor total variation restoration module, and a total variation fast image restoration module; wherein the support vector machine classifier module is used to divide the input data set image into images with stains and images with long-term degradation; the two types of images that have been divided are subjected to initial denoising using the total variation smoothing module, and the two different types of images after the initial denoising of the module are processed using the weighted factor total variation restoration module and the total variation fast image restoration module, and finally a conclusion is drawn by comparing the performance of the two images under the two restoration modules.

[0017] Furthermore, the essence of the overall variational smoothing denoising module is to regard the image as a piecewise smooth function and to establish a model for the image in a bounded space; when implementing this method, the partial differential equation must be digitized, converted into a difference equation, and the partial differential equation is solved through iterative operations. The biggest advantage of TV is that it overcomes the shortcomings of linear filtering in smoothing the edges of the image while suppressing noise. It does not change or sharpen the edges of the image while suppressing noise. Using this feature, the TV model can be used to retain the edge information of the image during the image repair process; at the same time, the total variation of the image can be minimized to achieve repair of the degraded image. The steps are:

[0018] Step 1: Read the gray value u of each pixel of the noisy image after classification by the SVM classifier (1) 0 ;

[0019] Step 2: Initialize parameter iteration coefficient n = 0, time step dt = 0.125, image initial gray value u 0 =u 0 , image divergence divp = 0;

[0020] Step 3: Determine the number of iterations. When n < the maximum number of iterations, n = n + 1, and calculate the next step u according to the discrete iteration formula. n The formula is as follows:

[0021]

[0022] Where: i, j represents the gray value of the point with coordinates (i, j) in image u, represents the n+1th iteration value, λ is the Lagrange multiplier, and dt represents the time step;

[0023] Step 4: Calculate the value of the image scatter term divp, the formula is as follows:

[0024]

[0025] Where: u x represents the first-order partial derivative of image u in the x direction; u y represents the first-order partial derivative of image u in the y direction; u xx represents the second-order partial derivative of image u in the x direction; u yy represents the second-order partial derivative of image u in the y direction; u xy Represents the second-order partial derivative of image u in the x direction first and then in the y direction;

[0026] Step 5: End the iteration, and the last iteration is the denoised image.

[0027] Furthermore, the key idea of ​​the weighted factor overall variational repair module is that the problem of finding the extreme value of the functional under constraints has morphological invariance and is suitable for repairing small areas such as scratches. Its characteristic is that the weight factor is added so that the pixel value in the central area of ​​the image has a greater influence on the next iteration result than the pixels in its eight-field area. The basic idea is to multiply the pixel data in the eight fields of the image repair point by the weight factor to reduce the amount of information provided by the eight fields of the repair point and increase the amount of information in the center of the repair point. The data of the former item is multiplied by a smaller weight value, and the data of the latter item is multiplied by a larger weight value, indicating that the data of the repair center point is given more attention, and the role of the data in its eight fields is gradually weakened, thereby shortening the repair time. The steps are:

[0028] Step 1: Read the initial grayscale value of each pixel in the image after denoising by the global variational smoothing denoising module and its mask information;

[0029] Step 2: Calculate the first-order derivative function and the modulus of its gradient for the pixels in the mask area of ​​the image;

[0030] Step 3: Determine whether the image pixel is outside the area to be repaired. If so, set λ=1, otherwise set λ=0. The formula with λ is:

[0031] Cost function:

[0032] J λ (u) The Euler-Lagrange equation that satisfies the minimum value is:

[0033]

[0034] Where: D is the area to be repaired of image u, E is an image neighborhood surrounding D in the information-intact area of ​​image u, and u 0 is the initial value of the image, and u represents the restored image value;

[0035] Step 4: Calculate h for pixels in the mask area of ​​the image OP 、h OO Get the new pixel value and save it in the new image. The formula is as follows:

[0036]

[0037]

[0038]

[0039] Where: w P The inverse of the gradient at a point adjacent to the point O to be repaired is calculated according to the formula;

[0040] Step 5: Determine the number of iterations. If it is not satisfied, repeat step 2. If the maximum number of iterations has been met, end the iteration. The last iteration is the repaired image. The iteration formula is as follows:

[0041]

[0042] Furthermore, the improved total variation (TV) fast image restoration module is an improved algorithm that intends to restore point O in one operation by increasing the neighborhood range and improving the weighting factor. Its feature is that it adds a noise reduction method based on anisotropic diffusion. If there is a degraded image with a more obvious step effect in the restoration area, the model can successfully overcome the step effect and the restoration of texture information appears more natural. The specific steps are as follows:

[0043] Step 1: Read the image after denoising in module (2) And determine the area to be repaired Ω and its repair δΩ.

[0044] Step 2: Calculate the repair priority of all points on δΩ based on the priority function P(p). The formula is as follows:

[0045] Priority: P(p) = C(p)D(p)

[0046] Trust Item:

[0047] Data items:

[0048]

[0049] Where: C(p) represents the effective information content in the sample block where point p is located, A data item representing the direction constraint of the equal illumination lines of image pixels, It reflects that the direction and intensity of the iso-illuminance line at point p are perpendicular to the gradient direction at point p; F(i,j) represents the pixel value of the point with coordinates (i,j) in the image; G i (i=0~n-1) represents the gradient vector of all known pixels in the sample block where the point is located, and n represents the number of known pixels in the sample block; G x (i, j) represents the gradient vector in the x direction at the point with coordinates (i, j) in the image; G y (i, j) represents the gradient vector in the y direction of the point with coordinates (i, j) in the image; t = 0 ~ n-1, n represents the number of known pixels in the sample block, Gx(t) and Gy(t) represent the gradient vectors along the x and y directions of all known pixels in the sample block where point p is located;

[0050] Step 3: Calculate and repair the point p with the highest priority on δΩ. The formula is:

[0051]

[0052] Where: point p is located in the area to be repaired, q is the sample block ψ p For a known point in the range of 1, w(p,q) = cos(p,q)dst(p,q), where cos(p,q) is the deviation angle factor, which ensures that the contribution of the known pixel point q close to the direction of the isoluminance line to point p is higher than the contribution of the known pixel point q far from point p in the direction of the isoluminance line; dst(p,q) represents the geometric distance factor, which reduces the contribution of the known pixel point q far from point p to point p;

[0053] Step 4: Calculate the priority of the points to be repaired in the four neighborhoods of point p, and update the repair δΩ:

[0054] Step 5: Repeat steps 3 and 4 until the δΩ position is repaired.

[0055] Step 6: Perform n (n < 10) iterations in the region Ω according to the iteration formula to alleviate the staircase effect. The formula is as follows:

[0056]

[0057] Preferably, the weighted factor overall variation repair module step 5 iteration formula sets the parameter μ weight factor to be Adding a weight factor makes the pixel value in the center of the image have a greater impact on the next iteration result than the pixels in its four areas. The basic idea is to multiply the pixel data in the four areas of the image repair point by the weight factor to reduce the amount of information provided by the four areas of the repair point and increase the amount of information in the center of the repair point. The data of the former item is multiplied by a smaller weight value, and the data of the latter item is multiplied by a larger weight value, which means that the data of the repair center point is more valued, and the role of the data in its four areas is gradually weakened, further shortening the repair time.

[0058] Preferably, the process of calculating and repairing the p point with the highest priority on δΩ in step 2 of the overall variation fast image restoration module is:

[0059] (1) The calculation form of the trust term C(p) is:

[0060]

[0061] The trust term represents the amount of valid information in the sample block where point p is located, ensuring that the more valid information contained in the neighborhood of point p, the higher the repair priority. M(q)=0, otherwise M(q)=1.

[0062] (2) The calculation form of the data item D(p) is:

[0063]

[0064] The data term ensures that p points with strong linear structure in the neighborhood are repaired first. It reflects the direction and intensity of the equal illumination lines at point p.

[0065] (3) The calculation form of the priority P(p) is:

[0066] P(p)=C(p)D(p)

[0067] (4) Find the gradient vectors of all known pixels in the sample block where point p is located and calculate the average value G p , the formula is as follows:

[0068]

[0069]

[0070] In the formula, i = 0 ~ n-1, n represents the number of known pixels in the sample block. Since the direction of the iso-illuminance line is perpendicular to the gradient direction, G can be p Rotate 90° to get

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] The present invention uses trust terms and iso-illuminance line intensity constraints to constrain the repair order and ensure the ability to maintain linear structures. Secondly, by introducing two weighting factors, geometric distance and deviation angle, the iterative equation of the TV model is improved to be the weighted sum of neighborhood information, so that the defect points can be repaired in one operation. Finally, a noise reduction method based on anisotropic diffusion is added to eliminate the step effect and improve the visual effect. The experimental results also show that the improved algorithm can achieve a similar repair effect as the TV model in repairing small area scratches, but is better than the TV model in repairing texture information and greatly improves the repair speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a diagram for dividing the area to be repaired and the intact area of ​​the present invention;

[0074] Figure 2 is the pixel O to be repaired and the neighborhood graph of the present invention;

[0075] Figure 3 A model diagram for designing weighted factors of the present invention;

[0076] Figure 4 It is a flow chart of the overall variational (TV) smoothing model of the present invention;

[0077] Figure 5This is a flowchart of the global variation (TV) image restoration of the present invention;

[0078] Figure 6 The flowchart of fast image restoration of the improved TV model of the present invention;

[0079] Figure 7 This is the control method of the automatic book turning machine of the present invention;

[0080] Figure 8 This is a block diagram of the gesture control system of the automatic book turning machine of the present invention;

[0081] Fig. 9 This is a control flow chart of the automatic book turning machine of the present invention;

[0082] Fig.10 This is the workflow of the Relief-MRMR-SVM hybrid feature selection method of the present invention;

[0083] Fig.11 This is a structural diagram of the unsharp mask of the present invention.

[0084] Fig.12 It is the overall workflow diagram of the present invention. DETAILED DESCRIPTION

[0085] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0086] 1. Place the books and pictures to be restored on the desk, and use the intelligent book-turning machine to automatically turn the pages and change the books of degraded document images that need to be restored. At the same time, when the photo entry software determines the pictures to be restored through the human-computer interaction interface, the image processing computer drives the camera to capture the pictures, and automatically enters the pictures to be restored into the computer for display, realizing automatic collection and processing of pictures.

[0087] The intelligent book turning machine consists of a book placing table, a book turning rod and a book pressing mechanism, which are controlled by a motor and a steering gear respectively. It can realize functions such as lifting or lowering the book pressing, turning pages left and right, and recording the number of pages turned according to control instructions of different modes, which provides convenience for the invention to realize automatic input of degraded document images.

[0088] The intelligent book turning machine can accept control instructions in various modes to realize its book turning function, such as Figure 7 As shown, the user can connect a button controller through a serial port data cable to achieve button control; similarly, after connecting to a touchpad, touch control can be achieved; connecting to a microphone can control it through voice; in addition, there are gesture control and facial expression control.

[0089] Gesture control system framework for intelligent book flipping machine Figure 8 As shown in the figure, the system captures the user's gesture command data through Kinect, and selects the corresponding control command to be transmitted to the intelligent book turning machine according to the analysis results of dynamic and static gestures, so as to achieve the purpose of turning the book. Among them, the dynamic command controls the page turning direction of the automatic book turning machine, and the static command controls the number of pages turned by the automatic book turning machine.

[0090] All control commands are received by Kinect, which controls the robot to turn the pages of a book by acquiring human skeleton information and depth image information, as well as dynamic and static gestures. The specific steps are as follows: Fig. 9 As shown:

[0091] (1) Set the OK gesture as the start signal for controlling the book turning;

[0092] (2) Detect whether there is a user's gesture control command within 10 seconds. If there is no control command, the algorithm ends; if there is, go to step (3);

[0093] (3) Check whether the user makes a static gesture of End. If so, the algorithm ends; if not, go to step (4);

[0094] (4) After the user swings his arm to give an instruction on the direction of turning the book, it is determined whether the instruction is given by the left hand or the right hand, and the swinging direction is determined, thereby determining the direction of turning the book;

[0095] (5) After the user gives a static gesture command, the system recognizes and controls the page turning;

[0096] (6) Return to step (2).

[0097] Second, a large amount of effective information is extracted from the pictures to be repaired after being entered into the computer, and a data classification scheme based on support vector machine is proposed. The required data is divided into two categories, pictures with stains and blurred pictures. Visual and mathematical methods are used to quantify and identify features, and preliminary analysis is performed on the images of documents with stains, blur and degradation.

[0098] Support vector machine (SVM) is a kind of machine based on VC dimension theory and structural risk minimization principle. It has a good effect on data classification. Compared with the traditional learning method based on structural risk minimization, it has the advantages of solving nonlinear, regression and high-dimensional pattern recognition, and does not fall into the local optimal problem. Especially in dealing with small sample decision problems, it can make full use of limited data information to mine as much classification information as possible. The model is as follows: Assume that there is a sample (x 1 ,y 1 ), (x 2 ,y 2),……,(x n ,y n ), x∈R d , y∈{1,―1}, n is the total number of samples, d is the data dimension. The ultimate goal of classification is to construct a classification decision function:

[0099] f(x)=sgn(g(x)) (1)

[0100] In formula (1), sgn is the sign function, and g(x) determines whether the classifier is a linear classifier or a nonlinear classifier. If it is a linear function in the above formula, it is called a linear support vector classifier.

[0101] The ultimate goal of the support vector classification machine is to find a minimized function The optimization problem to be solved can be expressed as:

[0102]

[0103] Formula (2) is the basic expression of the support vector classifier, D = {(x 1 ,y 1 ),(x 2 ,y 2 ),...(x m ,y m )},y i ∈{-1,1} is a given training sample set, where w=(w 1 ;w 2 ;...;w d ) is the normal vector, which determines the direction of the hyperplane. b is the displacement term, which determines the distance between the hyperplane and the origin. Since the parameters w and b in the constraints are linear and the objective function is a quadratic function, the optimization problem to be solved becomes solving the Lagrangian operator:

[0104]

[0105] (L(w,b,α) is the Lagrangian function for finding the optimal solution, α i As parameters, is the objective function (wx i +b)≥1 is an additional condition).

[0106] After taking the derivative of the parameters w and b in the above formula and setting their partial derivatives equal to zero, we can get

[0107]

[0108] From the above formula we can get:

[0109]

[0110] According to Lagrange's optimization method and duality theory, the above problem can be transformed into

[0111]

[0112] In the above formula, α i It represents the Lagrangian operator corresponding to the i-th sample, and the parameter b is calculated as follows:

[0113]

[0114] Where: N NSV ={i|α i >0,i=1,2,...,m} is the subscript set of all support vector machines.

[0115] The final optimal classification function is:

[0116]

[0117] Although the optimization problem becomes easier to solve through dual transformation, it is inevitable that there are a few samples that cannot be distinguished in advance. i ≥0(i=1,2,…,n) to cover a certain error range to classify these samples. So the constraint condition becomes

[0118] y i (wx i +b)≥1-l i ,i=1,2,...n (9)

[0119] The optimization problem is transformed into:

[0120]

[0121] According to the Lagrangian optimization method, it is transformed into a dual Lagrangian operator:

[0122]

[0123] In the above formula, C is the penalty factor, which controls the penalty intensity of misclassified samples. The larger the C value, the heavier the penalty for the wrong sample.

[0124] However, in the application, the data we need to process cannot be separated linearly. To improve these problems, the deep learning target detection algorithm and image classification method are used to realize the automatic classification of degraded images in multi-target images. In terms of feature extraction, the LBP (Local Binary Patterns) local texture features are introduced, and combined with the GLCM (Gray Level Co-occurrence Matrix) global features to better express the image texture; in terms of feature selection, a hybrid feature selection method combining Relief, MRMR (Max-Relevance and Min-Redundancy) and SVM is proposed. Multiple experiments were conducted on three degraded document image datasets to verify the effectiveness of this method.

[0125] The Relief algorithm is a multivariate filtering feature selection method. The process of feature selection for binary classification is as follows: ① Randomly sample a sample M from the training set, and find the nearest neighbor sample S in the sample set of the same type as M, and find the nearest neighbor sample D in the sample set of different types from M; ② Update the feature weight according to the rule: if the distance between M and S in a certain feature is less than the distance between M and D, it means that the feature is useful for distinguishing the nearest neighbor samples of the same and different types, then increase the feature weight, otherwise reduce it; ③ Repeat the above process m times to obtain the average weight of each feature, the smaller the value, the weaker the distinguishing ability of the feature.

[0126] The MRMR algorithm is also a filtering feature selection method, which realizes feature selection by maximizing the correlation between features and categories in the feature space and minimizing the redundancy between features, which is expressed by equations (12) to (14).

[0127]

[0128]

[0129]

[0130] In the formula, I(x,y) represents the mutual information function; S represents the degraded image feature set, x i ,x j represents each feature; c represents the target category. Formula (12) calculates the correlation between the feature and the target category, formula (13) calculates the redundancy between features, and formula (14) is the objective function.

[0131] MRMR is a heuristic greedy algorithm that cannot guarantee the global optimality of feature subsets. To solve this problem, the integrated MRMR proposed by NICOLAS et al. can be used.

[0132] Relief uses the relationship between features and categories to assign feature weights without considering the correlation between features. Integrated MRMR considers both the correlation between features and categories and the redundancy between features, but each selected feature contributes evenly to the classifier. Combining the two algorithms can effectively make up for each other's shortcomings and obtain a better feature subset. However, both algorithms are filtering methods with low classification accuracy. Therefore, they need to be combined with encapsulation methods to use classifiers to evaluate feature subsets to obtain the optimal feature subset and improve classification accuracy. Combining the advantages and disadvantages of the above algorithms and the two feature selection methods, a hybrid feature selection method combining Relief-MRMR-SVM is proposed. The workflow is as follows: Fig.10 shown.

[0133] The image data is divided into three categories using the above improved method. They are clear pictures, pictures with stains and blurred pictures. The above classified pictures are repaired using a fast image repair method based on the improved TV model. By adding repair order, increasing the field range, improving weighting factors and other key technologies, the defect point can be repaired in one operation.

[0134] The essence of the total variation (TV) smoothing model is to regard the image as a piecewise smooth function and to establish a model for the image in a bounded space. When implementing this method, the partial differential equation should be digitized and converted into a difference equation. The partial differential equation is solved through iterative operations. It was first used to reduce the noise of images contaminated by noise. By establishing a unified repair model based on the principle of energy minimization and applying it to the field of image repair, the model is called the total variation model because the solution of the energy function is transformed using the variational principle. Since this model can extend the edge, it is very suitable for repairing images with small damaged areas. The TV model relies on the collective features of the image when repairing, has morphological invariance, and can repair natural images well.

[0135] Let Ω be R n In the bounded open subset, u is a locally integrable function, and its total variation is defined as

[0136]

[0137] Let u 0 is the observed image, so u 0 =u+n, the total variational denoising energy functional of the image is defined as

[0138]

[0139] Where λ is the Lagrange multiplier. From the variational method k, we know that the necessary condition for the energy functional to achieve a minimum value is to satisfy the Euler-Lagrange equation, that is,

[0140]

[0141] By gradient descent method, we get TV smoothing model

[0142]

[0143] Regularizing |▽u| yields the regularized model

[0144]

[0145] in, is the gradient operator, yes The regularization term is used to reduce the degradation in flat areas. After converting the global coordinates to the local coordinate system at any pixel in the image, the equation can be decomposed into the edge direction and the two directions orthogonal to the edge. The coefficients of each direction after decomposition control the diffusion intensity in that direction. The diffusion equation is essentially a nonlinear anisotropic diffusion equation, and its diffusion operator only operates along the image gradient. The diffusion coefficient is There is no diffusion in the gradient direction, so the details of the image are preserved. One advantage of the TV minimization method is that it allows discontinuities in the image, that is, it preserves the edges. However, both theoretically and numerically, it is shown that the TV norm restores the smooth area into a "piecewise" continuous area, that is, the gradient effect.

[0146] Use u ij Represents the grayscale value of the pixel in image u (width M, height N), u ij n represents the value of the nth iteration, dt is the time step, and the difference of the derivatives of each party is as follows:

[0147]

[0148] The formula for calculating divergence is:

[0149]

[0150] The discrete iteration formula is

[0151]

[0152] There are many choices for the regularization parameter λ. Here we choose the inverse of the image gradient threshold. The corresponding steps are as follows:

[0153] (1) Read the gray value u of each pixel in the noisy image 0 ;

[0154] (2) Initialization parameters n = 0, dt = 0.125, u 0 =u 0 , divp=0;

[0155] (3) When n < the maximum number of iterations, repeat the following operations: ① n = n + 1, calculate the next step u according to the discrete iteration formula n ; ② Calculate the value of the divergence term;

[0156] (4) End the iteration, and the last iteration is the denoised image.

[0157] The key idea of ​​the weighted factor overall variation repair module is that the problem of finding the extreme value of the functional under constraints is morphologically invariant and suitable for repairing small areas such as scratches. Figure 1 As shown in the figure, D represents the area to be repaired, and E is an image neighborhood surrounding D in the area with intact information. Define u 0 is the initial value of the image, u represents the restored image value, and the restoration process of the TV model is the process of energy propagating from E to D.

[0158] The energy function on E∪D is defined as:

[0159]

[0160] Where u(x,y) represents the image, r is a non-negative real function. This formula is to ensure that the area to be repaired and its boundary are as smooth as possible. At the same time, in order to achieve good robustness to noise during the repair process, the following formula should also be satisfied:

[0161]

[0162] Where A(E) represents the area of ​​the domain E, and the initial image value u in E 0 Contaminated by Gaussian white noise, σ represents the standard deviation of white noise. In order to meet the requirements of patching edges and smooth the patched area as much as possible, an appropriate γ function must be selected. is an impact function δ. If we want to restore the damaged step edge, we need the energy function R to be finite. Therefore:

[0163]

[0164] Expand the function r(s) to become r(s)=s a +f(s), f(s) represents the low-order expansion term. According to formula (16), when s→∞, r(s) is finite, so a≤1 is deduced. Here we take a=1, that is, Then we can get the classic image restoration equation based on TV model.

[0165] According to the Langrange multiplier method, the above constrained equations (23) and (24) are transformed into unconstrained extreme value problems: To simplify the resulting equation, the multiplier is taken as Get the new cost function:

[0166]

[0167] The first term in the above formula is the total variation of the output image, which ensures the smoothness of the area to be repaired; the second term is the square error between the output image and the input image, called the fidelity term, which tries to maintain the similarity of the denoised image in the intact neighborhood. According to the variational principle, we know that J λ (u) The Euler-Lagrange equation that satisfies the minimum value is:

[0168]

[0169] In general, since the Euler-Lagrange equation is a nonlinear partial differential equation, for the original image u 0 =u 0 (x,y) is difficult to express analytically, so approximate numerical difference equations are often used instead in actual solutions.

[0170] Unsharp masking (UM) is a commonly used method for image edge and detail enhancement. Unsharp masking is to obtain a blurred image by pre-processing the original image with blur such as spatial smoothing filtering. The blurred image is subtracted from the original image to obtain the high-frequency information in the image, and then multiplied by a correction factor to adjust the obtained high-frequency information. The corrected high-frequency information is superimposed on the original image, so as to achieve the effect of enhancing the image contour and sharpening the image. Infrared images have large spatial correlation and blurred visual effects due to the long wavelength of infrared light, which is susceptible to atmospheric attenuation and long propagation distance. Therefore, the details of the image can be enhanced by the unsharp masking method. The structural block diagram is shown below. Fig.11 .

[0171] The formula of the unsharp mask algorithm is expressed as:

[0172] g(x,y)=f(x,y)+k*f mask (x,y) (28)

[0173] Among them, f(x,y) represents the original image, and f(x,y) is obtained by spatial smoothing filtering. Unsharp Mask -

[0174]

[0175] The mask part is the high-frequency information of the image, including 1) the detail layer in the image, 2) the existing noise, and 3) the overshoot and undershoot caused by the blurring of sharp edges. Commonly used spatial smoothing filters include Gaussian filtering, mean filtering, etc. k (k ≥ 0) is the high-frequency information gain coefficient, which controls the degree of detail enhancement. The larger the k, the more obvious the details of g (x, y); the smaller the k, the closer g (x, y) is to the original image. Unsharp masking can be divided into two types: linear and nonlinear. Linear unsharp masking means that the gain coefficient k of the high-frequency component of the image is a constant, while nonlinear unsharp masking means that the gain coefficient k is no longer a constant, but a variable k'(x, y) related to the image information. Therefore, nonlinear unsharp masking is also an adaptive method.

[0176] Based on the gradient descent method, a half-point format is used for discretization. Figure 2 The point to be repaired is O, and the 3×3 neighborhood of point O is considered: E, N, W, S are the four neighboring points of O. It is denoted as Λ = {N, S, E, W}, M = {n, s, e, w} is an imaginary half-pixel neighborhood point. The purpose of introducing half-pixels is to stabilize numerical calculations and reduce errors. Then the divergence function div can be approximately discretized by the central difference method as follows:

[0177]

[0178] In the formula, h is the step size. Based on the consideration of image discretization calculation, h=1 is generally taken.

[0179] The derivative value and gradient at the middle pixel {n, s, e, w} in equation (29) need to be further calculated. Take the half-pixel e as an example:

[0180]

[0181]

[0182] The above uses the central difference method to estimate the partial derivative and gradient in the x direction at e. Its advantage is that its value remains unchanged after rotating 90°, so the partial derivative and gradient at w, n, and s can be estimated in the same way. It can be known that the solution of the divergence function is

[0183]

[0184] Substituting the above equation into equation (27), we can obtain the discretized equation:

[0185]

[0186] To simplify the repair model equation, define

[0187]

[0188] Transforming equation (33) and substituting equation (34) into it yields

[0189]

[0190] Because the above formula satisfies The spectral radius of the matrix is ​​less than or equal to 1, so equation (35) is stable.

[0191] According to the conclusions of multiple tests, it is concluded that the weighting factor μ is related to the degree of image degradation and presents an inverse trend. When the image degradation is too serious, μ takes a larger value, and vice versa when the degree of degradation is light, μ takes a smaller value. The weighting factor of parameter μ is set to The addition of weight factors makes the pixel values ​​in the center of the image have a greater impact on the next iteration result than the pixels in its eight areas. The basic idea is to multiply the pixel data in the eight areas of the image repair point by the weight factor to reduce the amount of information provided by the eight areas of the repair point and increase the amount of information in the center of the repair point. The data of the former item is multiplied by a smaller weight value, and the data of the latter item is multiplied by a larger weight value, which means that the data of the repair center point is given more attention, and the role of the data in its four areas is gradually weakened.

[0192] Using the Gauss-Jacobi iterative algorithm, the image restoration value is obtained by the following iterative form:

[0193]

[0194] According to formula (36), the steps of the algorithm are as follows:

[0195] (1) Read in image and mask information;

[0196] (2) For each pixel in the image mask area, execute steps (3), (4), and (5);

[0197] (3) Calculate the first-order derivative and gradient modulus of the surrounding intermediate pixels;

[0198] (4) If the pixel is outside the area to be repaired, set λ = 1, otherwise set λ = 0;

[0199] (5) Calculate h OP Such as and h OO , get the new pixel value and save it in the new image;

[0200] (6) Replace the old image with the new image and go to step (2) until the number of iterations is completed.

[0201] By observing the repair value result (36) obtained by the above TV model algorithm iteration, it is found that λ(O)=0 in the defect area D, so the iteration equation becomes:

[0202]

[0203] It can be seen that the TV repair model repairs the damaged area D through weighted average operation, but because each operation only considers the influence of the gradient size of the 8 neighborhoods, the information obtained is limited, resulting in multiple iterations and loss of texture information. To address this problem, the following improved algorithm is proposed to repair point O in one operation by increasing the neighborhood range and improving the weighting factor.

[0204] like Figure 3 As shown, point p is located in the area to be repaired, and q is the sample block ψ p First, based on the general principle of image restoration, when repairing the current unknown point p, the closer the known point q is to it, the more effective information it contains. Therefore, the geometric distance between the two points will be an important factor in constraining the effect of q on p. Secondly, the weighting coefficient of the TV restoration model Only the magnitude of the gradient is considered, but the gradient is actually a vector that includes both magnitude and direction, and the influence of the gradient direction must be considered. Figure 3 In the figure, the q-point iso-illuminance line (a line connecting pixels with the same gray level) The direction is perpendicular to the direction of the gradient, indicating the direction of the smallest gradient change at point q. Two points on the same iso-illuminance line have the same grayscale value. The direction of the iso-illuminance line is normalized by the gradient size. For point q, if The smaller the angle of the direction of point p, the greater the repair effect on point p. The smaller the angle of the direction of q deviates from the direction of point p, the smaller the repair effect on point p. Therefore, the deviation angle between the iso-illuminance line direction of point q and the direction of point p closely affects the effect of q on p.

[0205] Based on formula (37), a new restoration model is designed by introducing two weighting factors: geometric distance and deviation angle. The formula is as follows:

[0206]

[0207] in:

[0208] w(p,q)=cos(p,q)dst(p,q) (39)

[0209]

[0210] The deviation angle factor cos(p,q) ensures that the contribution of the known pixel q near the isoluminance line to point p is higher than the contribution of the known pixel far from point p. The geometric distance factor dst(p,q) reduces the contribution of the known pixel q far from point p to point p.

[0211] For the repair process that advances from the outside to the inside through neighborhood information and completes the repair of defective pixels with only one calculation, the correct execution of the repair order helps to improve the repair effect. Figure 3 Ω represents the area to be repaired, δΩ represents the boundary line of the area to be repaired, and ψ p is the sample block to be repaired centered at point p on δΩ, n p is the normal vector, The direction of the iso-illuminance line (a line connected by pixels with the same grayscale level) of point p is represented. The priority function P(p) in the sample block-based image restoration technology is used to determine the restoration order of the fast restoration model, ensuring that the point p with more effective information and stronger linear structure in the neighborhood has a higher restoration priority.

[0212] The repair process of the improved model is based on local operations, so there will inevitably be a corresponding staircase effect, which will affect the visual effect. The seam effect is similar to the noise in the image, which is caused by some "alternative points" that lead to visual discontinuity. In order to better meet the visual connectivity and naturalness, a new denoising method based on anisotropic diffusion is used to alleviate the staircase effect while retaining image detail information.

[0213] like Figure 2 As shown, let Λ={N,S,E,W}, T=(NW,SW,NE,SE). Taking the four neighborhoods of the central pixel O as an example, define If it is in the smooth area of ​​the image, the grayscale difference |u O -u P | should be small, they belong to the "same type" point with point O and have a high correlation. If point E in the neighborhood of point O belongs to the "other type", then |u O -u P | should be large, and the correlation between point E and point O should be small. From the above analysis, we can see that the information of the central pixel is closely related to the neighboring pixels, and the influence of the neighboring points on the central point is inversely proportional to the grayscale difference between them. Therefore, in the process of denoising, the re-estimation of the central point pixel value only needs to consider the neighboring points that belong to the "same category" as it, and try to exclude the influence of "other" ones.

[0214] Therefore, for the central pixel O, considering the influence of the pixel information of its eight neighboring points, the influence function is defined as:

[0215]

[0216] Among them, n represents the current iteration number, g p It is the proportion of the grayscale difference between the central pixel and a point in the neighborhood to the grayscale difference between the central point and all the pixels in the neighborhood. Its expression is as follows:

[0217]

[0218] Furthermore, based on the anisotropic partial differential operator, the influence function defined above is introduced to obtain a new diffusion function:

[0219]

[0220] For the case where the neighborhood of the central pixel is selected as 8 pixels in the horizontal, vertical and diagonal directions, the discrete form of the diffusion function of the denoising method is as follows:

[0221]

[0222] During the experiment, q=0.4, the method fully considers the neighborhood information when re-estimates the central pixel. It is simple and easy to implement, overcomes the difficulty in selecting the parameter k in the PM model, and can retain detail information while reducing noise. This method can effectively alleviate the staircase effect and improve the visual effect of the repaired image.

[0223] The above algorithm flow is as follows:

[0224] (1) Read the image to be repaired, determine the area to be repaired Ω and its repair δΩ.

[0225] (2) Calculate the repair priority of all points on δΩ based on the priority function P(p).

[0226] (3) Repair the point p with the highest priority on δΩ according to equation (38).

[0227] (4) Calculate the priority of the points to be repaired in the four neighborhoods of point p and update the repair δΩ.

[0228] (5) Repeat (3) (4) until the δΩ position is repaired.

[0229] (6) Perform n (n < 10) iterations according to formula (44) in the region Ω to alleviate the staircase effect.

[0230] The experiment was conducted on a 256×256 degraded image. The performance was verified by repairing the degraded defective area. After the stained image and the long-term degraded image were smoothed and denoised by the TV model, the image was repaired by the fast algorithm based on the weighted factor TV model and the fast algorithm based on the improved TV model, and the repair results of the two were compared. The repair neighborhood is generally 9×9, which can be dynamically adjusted according to the size of the defective area. The results show that for degraded images with obvious step effects in the repair area, the fast algorithm of the improved TV model successfully overcomes the step effect, which better meets the requirements of visual psychology, indicating that the addition of a denoising method based on anisotropic diffusion can effectively eliminate the step effect. Comparing the repair results of the original TV repair model and the two improved repair model methods with the addition of weighted factors, the two can achieve similar repair effects in the repair of small scratches, but the multiple iterations of the original TV model lead to block effects and the inability to repair texture information. On the contrary, the fast algorithm of the improved TV model appears more natural in the repair of texture information. In the repair of larger degraded defective areas, compared with the improved method's ability to partially restore texture information, the TV model's block effect and inability to repair texture information are more prominent, but it has better linear structure preservation capabilities.

[0231] In order to objectively evaluate the performance of the improved algorithm, based on human visual psychology, three evaluation indicators, namely mean square error (MSE), peak signal-to-noise ratio (PSNR) and restoration time (t), were selected. The experimental results show that the MSE value and PSNR show an inversely proportional trend. The smaller the MSE value, the larger the PSNR value, indicating that the restoration result is more similar to the original complete image. Through performance analysis of images with stains, it can be seen that the MSE value of 192.04 of such images after using the fast image restoration based on the improved TV model is significantly lower than the MSE value of 283.741 after using the TV restoration model based on the weighting factor. The MSE value of 271.902 of another type of long-term degraded images after using the TV restoration model based on the weighting factor is significantly lower than the MSE value of 416.627 after using the fast image restoration based on the improved TV model, and the restoration time is greatly shortened from 5031ms to 141ms. The following conclusions can be drawn:

[0232] (1) The improved fast algorithm can achieve a similar repair effect as the original TV repair model in repairing small area scratches, and is better than the TV model in repairing texture information. The improvement is effective.

[0233] (2) For images with stains, fast image restoration based on the improved TV model can be used, while long-term degraded images can be restored using the TV model based on weighted factors.

[0234] (3) The TV model algorithm based on weighted factors is significantly better than the improved fast algorithm in terms of running time and the original TV repair model, but its ability to maintain linear structure is weaker than that of the TV repair model.

Claims

1. A degraded image restoration method based on intelligent page turning and accelerated TV-PDE, It is characterized in that The specific steps are as follows: (1) Place the image book to be restored on the intelligent book turning machine, and use the intelligent book turning machine to automatically turn pages and change books of the degraded document image book to be restored; at the same time, when the photo input software determines the image to be restored through the human-computer interaction interface, the image processing computer drives the camera to capture the image, and automatically inputs the taken image to be restored into the computer for display, thereby realizing automatic image acquisition processing; (2) Extract effective information from the images to be restored after they are entered into the computer, use visual and mathematical methods to quantify and identify the features of stained and blurred images, and analyze and restore stained, blurred and degraded document images; (3) The analysis and restoration method in step (2) specifically includes a support vector machine classifier module, a global variational smoothing denoising module, a weighted factor global variational restoration module, and a global variational fast image restoration module; wherein the support vector machine classifier module is used to divide the input data set image into images with stains and images with long-term degradation; the two types of images that have been divided are subjected to initial denoising using the global variational smoothing module, and the two different types of images after the initial denoising of the module are processed using the weighted factor global variational restoration module and the global variational fast image restoration module, and finally a conclusion is drawn by comparing the performance of the two images under the two restoration modules; The overall variational smoothing denoising module has the following specific operation steps: Step 1: Read the grayscale value u of each pixel of the noisy image after the support vector machine classifier module SVM classifier classification processing 0 ; Step 2: Initialize parameter iteration coefficient n = 0, time step dt = 0.125, image initial gray value u 0 =u 0 , image divergence divp = 0; Step 3: Determine the number of iterations. When n < the maximum number of iterations, n = n + 1, and calculate the next step u according to the discrete iteration formula. n The formula is as follows: Where: i, j represents the gray value of the point with coordinates (i, j) in image u, represents the n+1th iteration value, λ is the Lagrange multiplier, and dt represents the time step; Step 4: Calculate the value of the image scatter term divp, the formula is as follows: Where: u x represents the first-order partial derivative of image u in the x direction; u y represents the first-order partial derivative of image u in the y direction; u xx represents the second-order partial derivative of image u in the x direction; u yy represents the second-order partial derivative of image u in the y direction; u xy Represents the second-order partial derivative of image u in the x direction first and then in the y direction; Step 5: End the iteration, and the last iteration is the denoised image; The weighted factor overall variation repair module has the following specific operation steps: Step 1: Read the initial grayscale value of each pixel in the image after denoising by the global variational smoothing denoising module and its mask information; Step 2: Calculate the first-order derivative function and the modulus of its gradient for the pixels in the mask area of ​​the image; Step 3: Determine whether the image pixel is outside the area to be repaired. If so, set λ = 1, otherwise set λ = 0. The formula with λ is: Cost function: J λ (u) The Euler-Lagrange equation that satisfies the minimum value is: Where: D is the area to be repaired of image u, E is an image neighborhood surrounding D in the information-intact area of ​​image u, and u 0 is the initial value of the image, and u represents the restored image value; Step 4: Calculate h for pixels in the mask area of ​​the image OP 、h OO Get the new pixel value and save it in the new image. The formula is as follows: Where: w P The inverse of the gradient at a point adjacent to the point O to be repaired is calculated according to the formula; Step 5: Determine the number of iterations. If it is not satisfied, repeat the weighted factor overall variation restoration module step 2. If the maximum number of iterations has been satisfied, end the iteration. The last iteration is the restored image. The iteration formula is as follows: Set the parameter μ weight factor to The overall variation fast image restoration module has the following specific operation steps: Step 1: Read the initial grayscale value of each pixel in the image after denoising by the global variational smoothing denoising module And determine the area to be repaired Ω and its repair δΩ; Step 2: Calculate the repair priority of all points on δΩ based on the priority function P(p). The formula is as follows: Priority: Trust Item: Data items: Where: C(p) represents the effective information content in the sample block where point p is located, D(p) represents the data item of the direction constraint of the equal illumination line of the image pixel, F(i,j) represents the pixel value of the point with coordinates (i,j) in the image; G i (i=0~n-1) represents the gradient vector of all known pixels in the sample block where the point is located, and n represents the number of known pixels in the sample block; G x (i, j) represents the gradient vector in the x direction at the point with coordinates (i, j) in the image; G y (i, j) represents the gradient vector in the y direction of the point with coordinates (i, j) in the image; t = 0 ~ n-1, n represents the number of known pixels in the sample block, Gx (t) and Gy (t) represent the gradient vectors of all known pixels in the sample block where point p is located along the x and y directions respectively; the trust term represents the amount of valid information in the sample block where point p is located, ensuring that the more valid information is contained in the neighborhood of point p, the higher the repair priority is. M(q)=0, otherwise M(q)=1; The data term ensures that the p points with strong linear structure strength in the neighborhood are repaired first. It reflects the direction and intensity of the iso-illuminance line at point p; in the data term, i = 0 ~ n-1, n represents the number of known pixels in the sample block. Since the direction of the iso-illuminance line is perpendicular to the gradient direction, G can be p Rotate 90° to get Step 3: Calculate and repair the point p with the highest priority on δΩ. The formula is: Where: point p is located in the area to be repaired, q is the sample block ψ p For a known point in the , w(p,q)=cos(p,q)dst(p,q), where cos(p,q) is the deviation angle factor, which ensures that the contribution of the known pixel point q close to the direction of the iso-illuminance line to p is greater than the contribution of the known pixel point far from the direction of the iso-illuminance line; dst(p,q) represents the geometric distance factor, which reduces the effect of the known pixel point q far away from point p on point p; Step 4: Calculate the priority of the points to be repaired in the four neighborhoods of point p, and update the repair δΩ; Step 5: Repeat steps 3 and 4 of the overall variation fast image restoration module until the δΩ bit space is restored; Step 6: Perform n iterations in the region Ω according to the iteration formula, n < 10, to alleviate the staircase effect. The formula is as follows: Where: Represents the n+1th iteration value of the image, is the acceleration coefficient, and q=0.4 is the weighting factor.

Citation Information

Patent Citations

  • Improved image enhancement method based on block matching and recovery and combined trilateral steering filtering

    CN106485672A

  • Hole filling algorithm based on depth transition and overall depth item variability

    CN106998460A

  • Image restoration method based on low-rank tensor chain decomposition

    CN111598798A

  • Four-direction relative total variation image denoising method

    CN113112425A

  • Method for processing image and electronic device for the method

    US20150365624A1