CN114998632B - RVIN detection and removal method based on pixel clustering - Google Patents

Abstract

The application discloses a RVIN detection and removal method based on pixel clustering, which comprises the following steps: clustering and segmentation are carried out based on the gray distance similarity of the pixel points, and all pixels in the damaged image are classified into K classes; calculating LCI values of pixels, determining an area where the pixels are located based on the LCI values, wherein the area comprises a flat area and a detail area, obtaining an optimal detection threshold value of each type of pixels through iterative solution, and judging whether the pixels are noise pixels or not according to the LCI values of the pixels and the optimal detection threshold value; an LCI weighted mean filter and an edge direction filter are employed for noise pixels in the flat and detail regions, respectively, to recover pixels corrupted by random value impulse noise. The noise detector and the filter provided by the application have high robustness and generalization, and have remarkable effects in RVIN removal of natural images and medical images, and particularly have better effects on high noise level.

Description

A method for RVIN detection and removal based on pixel clustering

Technical field

The present invention relates to the field of image processing technology, and in particular to a method of detecting and removing RVIN based on pixel clustering.

Background technique

Digital images are easily interfered by noise during the generation and propagation process, which affects subsequent image processing. Image denoising has always been one of the underlying visual tasks that urgently needs to be solved. The image denoising model can be expressed as Y=X+B, where Y and X represent damaged images and clear images respectively, and B represents adaptive noise. The noise existing in natural images and medical images is mainly Gaussian noise and random impulse noise (random-valued impulse noise, RVIN). There are many algorithms for Gaussian noise removal and they perform well. When the image is contaminated by random impulse noise, only some pixels in the image are destroyed, and the new gray value of this part of the pixels is randomly between 0 and 255. Compared with Gaussian noise, salt-and-pepper noise and other types of noise, this random characteristic brings more trouble to noise removal.

Several filters for Gaussian noise removal such as Gaussian filter, mean filter and bilateral filter were tried to remove RVIN. These linear and nonlinear edge-preserving filtering methods are a compromise process that combines the spatial proximity and pixel value similarity of the image, taking into account spatial information and grayscale similarity to achieve the purpose of edge-preserving denoising, but for random The effect of impulse noise is not significant and tends to blur the repaired image. The median filter and its variant CWM sort the pixels in a local area according to the gray level, and take the median or weighted median of the gray level in the area as the gray level value of the current pixel. In contrast, they have a better effect on removing impulse noise, and at the same time overcome to a certain extent the problem of blurring image details processed by linear filters. Therefore, a new adaptive weighted median filter (ACWM) is proposed, which uses a learning method based on the least mean square (LMS) algorithm to obtain the center weight within each block, and then gradually applies noise filtering through multiple iterations program to obtain the optimal filtering effect. However, for images with many detailed textures such as points, lines, and spires, these median filters and their improved algorithms tend to remove normal pixels in details and textures as noise pixels, and cannot fundamentally solve image blur and details. Information loss problem.

The two-stage RVIN denoising algorithm based on noise detection and filtering first detects noise pixels in the damaged image and then removes them, which can effectively solve the problem of blurred images and loss of details in repaired images. Obviously, the denoising effect of this two-stage method is closely related to the accuracy of noise detection. In order to accurately screen out random impulse noise in damaged images, a noise detection method based on local statistical order absolute difference (ROAD) is proposed, which counts the grayscale difference between the central pixel in the local window and its neighbor pixels. to determine whether the pixel is noise. Inspired by the ROAD method, DONG proposed to use a logarithmic function to convert the ROAD value of the central pixel into a ROLD form to amplify the difference between the central pixel and its neighboring pixels, thereby improving the accuracy of impulse noise detection. Yu combined ROAD and ROLD and proposed a noise detection method based on rank ordered relative differences (RORD). However, these methods do not consider the statistical information of pixels within the window range, such as variance, and the prior knowledge of noise, such as noise level.

In fact, the level of noise has a great impact on the performance of denoising algorithms. In order to make full use of the prior knowledge of image noise level, a noise detection method based on local consensus index (LCI) is proposed. By calculating the LCI value of the central pixel, it measures its similarity with other pixels in the neighborhood, and then estimates the degree of similarity between it and other pixels in the neighborhood. The noise level of the damaged image is used to set an appropriate LCI threshold to filter out normal pixels and noise pixels. In order to improve the detection accuracy of noise, it obtained the calculation formulas of LCI threshold and image noise level through a large number of experiments and polynomial fitting. However, for some images with complex textures and severe damage, the LCI values of their pixels are generally low, which easily increases the number of misjudged pixels. The literature [Triple Threshold Statistical Detection filter for removing high density random-valued impulse noise in images] designs a three-threshold detection method of standard deviation, mean and quartile to solve the problem of image denoising under high noise levels. However, this multi-threshold detection method has little effect on improving noise detection while increasing the complexity of the algorithm, and cannot well distinguish between normal pixels and noise pixels in detail or texture areas. To solve this problem, Nadeem] designed a blur noise detector based on switching technology, which can well distinguish noise pixels and edge pixels in detail and texture areas. In the noise detection stage, by using appropriate thresholds, pixels in the image can be identified as normal pixels, noise pixels or candidate noise pixels, but this article does not specifically explain how to filter edge pixels from candidate noise. The literature [Liu] introduces in detail an order road difference (ROD-ROAD) and local image statistics minimum edge pixel difference (MEPD) methods to identify edge pixels from candidate noise to prevent edges from being misidentified as noise. These detection and filtering methods based on neighborhood pixels basically only consider the pixel gray value information within a limited window range, and do not consider the pixel distribution characteristics of the entire image, resulting in poor generalization performance of the algorithm. And when the image damage degree reaches 60% or even higher, these methods can easily misjudge normal pixels as noise pixels because there are more noises in the neighborhood than normal pixels.

Deep learning technology developed in recent years has also been widely used in image denoising. The proposed denoising method based on convolutional neural network is used to remove Gaussian noise. It is a powerful nonlinear mapping model in image processing. However, this model The flexibility is severely limited and is not applicable to RVIN. To solve this problem, a blind CNN model for RVIN denoising is proposed by using a flexible Noise Ratio Predictor (NRP) as an indicator. However, the training of these end-to-end neural network denoising models requires a lot of computing costs and has no obvious advantages over traditional denoising methods. The complexity of the model also makes it difficult to implement the algorithm.

In summary, although various image denoising algorithms are constantly being added, many noise detection methods that use manually set detection thresholds or are designed based on local window information do not have good generalization performance and cannot effectively handle high-noise level subjects. The image is damaged, and the noise pixels and edge pixels cannot be accurately distinguished. As a result, the details or edge information of the image are often lost during noise reduction.

Contents of the invention

In order to solve the above technical problems, the present invention proposes a method of RVIN detection and removal based on pixel clustering. In order to improve the generalization performance of the denoising algorithm and achieve rapid noise reduction while still retaining sufficient detailed information, the present invention The noise detector is based on the idea of grouping and clustering. It divides all pixels in the damaged image into several categories according to the characteristics of the pixels, and then identifies the noise of each group of pixels through adaptive thresholds; based on the noise detection results, the present invention proposes a Filters for partition decision-making use different filters for noise pixels in different areas to restore pixels damaged by random value impulse noise. A large number of experimental results show that the proposed method is suitable for RVIN in natural images or medical images, and is substantially better than other advanced RVIN filtering techniques in terms of visual and objective quality measurements.

In order to achieve the above objects, the technical solutions of the present invention are as follows:

A method of RVIN detection and removal based on pixel clustering, including the following steps:

Cluster segmentation is performed based on the grayscale distance similarity of pixels, and all pixels in the damaged image are divided into K categories;

Calculate the LCI value of the pixel and determine the area where the pixel is located based on the LCI value. The area includes flat areas and detailed areas, and then obtain the optimal detection threshold for each type of pixel through iterative solution. According to the LCI value of the pixel and the optimal detection The threshold determines whether the pixel is a noise pixel;

The LCI weighted mean filter and the edge direction filter are used respectively for the noisy pixels in the flat area and the detailed area to restore the pixels damaged by the random value impulse noise.

Preferably, the cluster segmentation also includes the following steps:

The image is smoothed, including median filtering and Gaussian filtering.

Preferably, the clustering is performed based on the gray distance similarity of pixels, and the clustering method adopts K-means clustering method, which specifically includes the following steps:

Find K cluster centers μ _k (k=1,...,K), and assign all pixels in the damaged image to the nearest cluster center so that each pixel has a one-dimensional distance from its corresponding cluster center _The sum _of squares of n=1,...,N, k=1,...,K), if the pixel x _n belongs to the kth cluster, then r _nk =1, otherwise it is 0. The following loss function can be defined:

It can be seen from the above formula that the initial value of the cluster center μ _k needs to be randomly fixed to find the attribution value r _nk of the pixel that minimizes the loss function J. Given the gray value of the pixel x _n and the cluster center μ _k , the loss Function J is a linear function of r _nk . Since x _n and x _n+1 are independent of each other, for each pixel point x _n , the point only needs to be assigned to the nearest cluster center, that is

Use r _nk obtained in formula (2) and bring it into formula (1) to find the cluster center μ _k . Given the value of r _nk , the loss function J is the quadratic function of μ _k . Let the derivative of J with respect to μ _k is 0, we can get

From the above formula, it can be deduced that the value of μ _k is μ _k is the mean gray level of the pixels belonging to this category.

Preferably, the clustering method adopts mean shift clustering method, density-based clustering method or maximum expectation clustering of Gaussian mixture model.

Preferably, the LCI value of the calculated pixel is calculated and obtained by calculating the same type of pixels in the neighborhood of the pixel.

Preferably, obtaining the optimal detection threshold for each type of pixel through iterative solution includes the following steps:

Traverse the detection threshold from 0 to 1, and calculate the objective function of the image denoising model. When the objective function is the smallest, the current detection threshold is the optimal detection threshold. The objective function is as follows:

Among them, y is any pixel of the image, i, j are the coordinates of y, and V(y) is called the TV norm, which is a regularization method with the goal of maintaining image edge information.

Preferably, the LCI weighted mean filter is as follows:

Among them, I' _x represents the gray value of ^the filtered noise pixel x, _Y represents the _pixel in _Ω value.

Preferably, the filter window of the LCI weighted mean filter is set to 5×5.

Preferably, the edge direction filter is used for noise pixels in detail areas to restore pixels damaged by random value impulse noise, which specifically includes the following steps:

For noise pixels that are determined to be detail areas, a detection frame is constructed with the noise pixel as the center;

Put the normal pixels in the row, column, left diagonal and right diagonal line centered on the noise pixel in the detection frame into the sets D _h , D _v , D _l and D _r respectively;

Calculate the standard deviation of the elements in the D _h , D _v , D _l , and D _r sets respectively, and select the direction represented by the set with the smallest standard deviation as the boundary filtering direction;

Arrange the grayscale values of the normal pixels in the boundary filtering direction in ascending or descending order, and select the median in the sequence as the new grayscale value of the central noise pixel.

Preferably, the filter window of the edge direction filter is set to 7×7.

Based on the above technical solution, the beneficial effects of the present invention are: the noise detector in the present invention is based on the idea of grouping and clustering, divides all pixels in the damaged image into several categories according to the characteristics of the pixels, and then identifies each group through adaptive thresholds Noise of pixels; based on the noise detection results, the present invention proposes a partition decision-making filter. Different filters are used for noise pixels in different areas to restore pixels damaged by random value impulse noise. The noise detector and filter proposed by the present invention have high robustness and generalization, and have achieved remarkable results in RVIN removal of natural images and medical images, especially at high noise levels.

Description of the drawings

Figure 1 is a flow chart of a method for RVIN detection and removal based on pixel clustering in one embodiment;

Figure 2 is a diagram of the clustering effect and pixel classification results of a method for RVIN detection and removal based on pixel clustering in one embodiment, wherein Figure 2(a) is a Lena image with a noise level of 50% when K=4 Clustering effect diagram; Figure 2(b) shows the pixel classification result of area A, in which the pixels marked in red, green and white indicate that they belong to different;

Figure 3 is a LENA damaged image noise detection situation in an RVIN detection and removal method based on pixel clustering in one embodiment, where Figure 3(a) is a 50% LENA damaged image; Figure 3(b) is Pixel detection situation diagram of area A in Figure 3(a);

Figure 4 is an example of six experimental test pictures of a method of RVIN detection and removal based on pixel clustering;

Figure 5 is a diagram of the detection effect of the noise detector on test images with different noise levels in one embodiment;

Figure 6 is a comparison diagram of 50% Lena image repair in an embodiment, wherein Figure 6(a) is the 50% Lena image; Figure 6(b) is the repaired Lena image;

Figure 7 is a ship image of 60\%RVIN and different restoration effects. Figure 7(a) is a ship image of 60\%RVIN; Figure 7(b) is the effect of using AFWMF to process Figure 7(a). Figure; Figure 7(c) is the rendering after using ASMF to process Figure 7(a); Figure 7(d) is the rendering after using BDND to process Figure 7(a); Figure 7(e) is the rendering after using DWM processing The rendering after 7(a); Figure 7(f) is the rendering after using EAIF to process Figure 7(a); Figure 7(g) is the rendering after using EBDND to process Figure 7(a); Figure 7( h) is the rendering after using FRDFN to process Figure 7(a); Figure 7(i) is the rendering after using ROR-NLM to process Figure 7(a); Figure 7(j) is using SBF to process Figure 7(a) ); Figure 7(k) is the rendering after using SDOOD to process Figure 7(a); Figure 7(1) is the rendering after using the method of this application to process Figure 7(a);

Figure 8 is a filtering effect diagram of a prostate noise image in an embodiment, wherein Figure 8(a) is a prostate image containing 30% RVIN; Figure 8(b) is a repaired prostate image;

Figure 9 is a filtering effect diagram of a head noise image in an embodiment, wherein Figure 9(a) is a head image containing 50% RVIN; Figure 9(b) is a repaired head image.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

As shown in Figure 1, this embodiment provides a method for RVIN detection and removal based on pixel clustering, which includes the following steps:

Step S101: Cluster segmentation is performed based on the grayscale distance similarity of pixel points, and all pixels in the damaged image are divided into K categories.

Among them, data clustering methods include mean shift clustering method, density-based clustering method (DBSCAN), Gaussian mixture model (GMM) expectation maximum (EM) clustering, etc. In this embodiment, the more commonly used K-means algorithm is used because it is simple and has fast convergence speed. K-means image clustering and segmentation, also known as K-means clustering, uses the principle of unsupervised learning to cluster pixels into several clusters. The specific principles are as follows:

Theorem 1: Given a set of data {x ₁ ,...,x _N } in D-dimensional Euclidean space, assuming that the number of clusters K is known, from the perspective of Euclidean space, the distance Closer points are grouped into a cluster, and points in different clusters are relatively far apart.

It can be seen from Theorem 1 that K cluster centers μ _k (k=1,...,K) need to be found, and all pixels in the damaged image are assigned to the nearest cluster center so that each pixel corresponds to The sum of squares of the one-dimensional distance between the cluster centers is the smallest, where the one-dimensional distance here refers to the grayscale difference between the two. We introduce a binary variable r _nk ∈ {0,1} to represent the belonging of a certain pixel x _n in the damaged image to cluster k (where n=1,...,N, k=1,... ,K), if the pixel x _n belongs to the kth cluster, then r _nk =1, otherwise it is 0. In this way, the following loss function can be defined:

It can be seen from Formula 1 that the initial value of the cluster center μ _k needs to be randomly fixed to find the attribution value r _nk of the pixel that minimizes the loss function J. Given the gray value of pixel point x _n and cluster center μ _k , the loss function J is a linear function of r _nk . Since x _n and x _n+1 are independent of each other, for each pixel point x _n , only The point needs to be assigned to the nearest cluster center, that is

Use r _nk obtained in formula (2) and bring it into formula (1) to find the cluster center μ _k . Given the value of r _nk , the loss function J is a quadratic function of μ _k . Let the derivative of J with respect to μ _k be 0, we can get

From the above formula, it can be deduced that the value of μ _k is μ _k is the mean gray level of the pixels belonging to this category.

After the above steps, the pixels of the image will be clustered based on similarity. The similarity of the pixels is calculated based on the one-dimensional distance between the surrounding pixels and the central cluster pixels, that is, the gray difference between the two. The minimum distance to the central cluster is found. pixels are grouped into one category. According to the value of K, the pixels of an image can be clustered into K categories. Through pixel clustering, complex image textures can be accurately separated logically, achieving class-to-class processing, and improving the ability to repair damaged images to a certain extent. Figure 2(a) shows the clustering effect of the lena image with an RVIN of 50% when the K value is 4. It can be seen from the figure that the pixels of the image are divided into four categories, and each category is displayed in a different color. At the same time, for pixels between classes, since the pixels with certain similarities have been classified in the clustering, when calculating its LCI, only the pixels of the same class in its neighborhood are used to calculate the LCI, that is Can. Figure 2(b) shows the pixel classification result of area A in Figure 2(a). For the central pixel x, there are only 8 pixels among the 24 pixels in its neighborhood (221,32,64,22,112,23 ,23,74) belong to the same category as it, then only these pixels are considered when calculating the LCI value of the center pixel x, so as to increase the prior knowledge of the edge. It should be pointed out that due to the poor anti-interference of the K-means algorithm, before clustering and segmenting the image, we will first perform two simple filtering methods, median filtering and Gaussian filtering, to weaken the effect of noise on clustering. Impact.

Step S102, calculate the LCI value of the pixel and determine the area where the pixel is located based on the LCI value. The area includes a flat area and a detailed area, and then obtain the optimal detection threshold for each type of pixel through iterative solution. According to the LCI value of the pixel and The optimal detection threshold determines whether the pixel is a noise pixel.

In this embodiment, the LCI value of the pixel is calculated. The specific calculation method is as shown in formulas (4)-(7):

In formula (4), Ω _x ⁰ is a 5×5 neighborhood centered on pixel x, y is any pixel in the neighborhood Ω _x ⁰ , u _x and u _y represent the grayscale values of pixels x and y, ( m,n) and (s,t) are the coordinates of pixel x and y respectively. σ _λ and σ _s are the parameters of the preset Gaussian kernel function respectively. Formula (4) indicates that the similarity θ(x,y) between two pixels is related to their distance and grayscale difference. It can be found from formula (5) that when x is a normal pixel, due to its high similarity with every other normal pixel in the neighborhood, the value of ζx will be larger, and vice versa. Therefore, by observing the ζx value of the central pixel x, the possibility that it is a normal pixel can be evaluated. In order to make the statistic ζx have better stability and discriminability, it is normalized through formulas (6) and (7) to limit it to the field [0,1], and finally the pixel x is obtained. LCI value. LCI represents the probability of whether a pixel is noise. The larger the LCI value of a pixel, the more likely it is to be a normal pixel. By setting an appropriate threshold, normal pixels and noise pixels in the entire damaged image can be filtered out. In order to obtain the best detection effect, LCI is first used to determine whether the pixel is in a flat area or a detailed area, and then different LCI thresholds are used to filter out the noise. This method is computationally simple and can quickly detect RVIN of damaged images without iteration.

After iterative calculation, the pixels of the image will be clustered into several different clusters based on the similarity of gray distances. Obviously, in different clusters, due to different grayscales and different regions of pixels, the corresponding noise detection threshold ranges will also be different. Therefore, it is now necessary to select optimal detection thresholds for different types of pixels. For the two-stage denoising algorithm, the better the noise detector, the better the filtering effect, which means that the optimal detection threshold corresponds to the best filtering effect. In other words, the selection of the optimal detection threshold of TLCI is actually a process of solving the image denoising model optimization problem.

Since the pixel LCI has been normalized and its value range is between [0,1], we can let the detection threshold TLCI traverse from 0 to 1 (assuming the step size is 0.1) while calculating the target of the image denoising model function, when the objective function is the smallest, the corresponding TLCI value is the optimal detection threshold. The total variation model is often used to solve optimization problems of image denoising. This model mainly relies on the gradient descent method to smooth the image. We hope to smooth the image inside the image so that the difference between adjacent pixels is small, and The outline (edge) of the image is not smoothed as much as possible, and image denoising is based on directional Laplacian regularization. Therefore, we use the characteristic that the image is a two-dimensional discrete signal to perform total variation on the image, as shown in formula (8):

where y is any pixel of the image, i, j are the coordinates of y. Since it is difficult to solve the total variation of formula (8), there is another commonly used definition of anisotropy in two-dimensional total variation:

In formula (9), V(y) is called the TV norm, which can be used as a regularization method to maintain image edge information. The TV value of the image is the same as the norm representation of the matrix. The anisotropic TV of the image is The norm is the L1 norm of the matrix, and the isotropic TV norm of the image is expressed in the same way as the L2 norm of the matrix. Therefore, when we use anisotropy V(y) as the objective function of the model, the repaired image V(y) value is the smallest, and the image repair effect is the best. Through this method, we can determine the optimal detection threshold TLCI of the noise detector. .

Because the pixels of the image are divided into K categories through the K-means method, and pixels belonging to the same category are divided into flat areas and complex areas, it is necessary to use an iterative method to select the optimal thresholds for 2K areas, and then restore the image while calculating its TV value (note that the regions here are relatively independent from each other). Experiments have found that as the detection threshold iterates from 0 to 1, the value of V(y) shows a trend of first decreasing and then increasing. Therefore, there is a threshold to minimize V(y), and the threshold at this time is what we call required optimal detection threshold.

Step S103: Use the LCI weighted mean filter and the edge direction filter respectively for the noise pixels in the flat area and the detail area to restore the pixels damaged by the random value impulse noise.

In this embodiment, different filters should be used in the filtering stage according to different areas where the noise pixels are located. A more robust partition decision filter is designed to remove RVIN instead of using the existing median or improved median filter. The proposed partition decision filter takes into account both the image characteristics and the area where the noise is located, and only selects pixels judged to be normal in the neighborhood of the central pixel to filter the central pixel, so it is more suitable for removing RVIN noise. .

For noise pixels judged to be in flat areas, they are repaired through the LCI weighted mean filter. The LCI weighted mean filter looks like this:

Among them, I' _x represents the gray value of ^the filtered noise pixel x, _Y represents the _pixel in _Ω value. The LCI value of a pixel is used as the weight of each pixel in the filter because LCI represents the probability that a pixel is a normal pixel. If the LCI value of a pixel is larger, it indicates that it is more likely to be a normal pixel, so it should be given more weight. Participate in the repair of center noise pixels. Considering that the pixel grayscale distribution in the flat area is smooth, if the filter window is set too large, it is easy to introduce pixels in the edge area, so the filter window of the LCI weighted mean filter is set to 5×5.

For noise pixels that are judged to be in detail areas (edge areas), the grayscale of pixels in its neighborhood changes drastically. However, due to the characteristics of the edge, there will always be a grayscale difference of pixels in a certain direction in this neighborhood. smaller. Therefore, we designed a median filter based on the minimum gradient difference. The specific process is as follows:

Step S131: For the noise pixels determined to be detail areas, construct a 7×7 detection frame centered on the noise pixels;

Step S132, put the normal pixels in the row, column, left diagonal and right diagonal line centered on the noise pixel in the detection frame into the set D _h , D _v , D _l and D _r respectively. middle;

Step S133, calculate the standard deviation of the elements in the D _h , D _v , D _l , and D _r sets respectively, and select the direction represented by the set with the smallest standard deviation as the boundary filtering direction;

Step S134: Arrange the grayscale values of the normal pixels in the boundary filtering direction in ascending or reverse order, and select the median in the sequence as the new grayscale value of the central noise pixel.

Select an edge area A from the 50% LENA damaged image in Figure 3(a), and the corresponding pixel grayscale distribution of this area is shown in Figure 3(b). The elements in the four-direction set of this area are D _h =[56,95,211], D _v =[106,107,99,85,72], D _l =[90,80,215], D _r =[147,117,67,54 ,38]. From the standard deviation, we can know that the direction of D _v is the boundary line. From the figure, we can also know that the direction of this boundary line is consistent with the direction of the boundary line calculated by us. Then after median filtering, the gray value of the central pixel is 99 in the D _v set, which is very close to the effective and correct data 91 of the pixel. It should be pointed out that if the window of the median filter is too small, there may be very few normal pixels in the edge direction, which causes the image to be prone to burrs and affects the filtering effect. Therefore, we will improve the median filter. The window is set to 7×7.

experiment

A large number of experiments have been done on standard natural images and medical images. The test images are shown in Figure 4. Except for the house image in Figure 4(c), which is 256×256, the other image sizes are 512×512.

1. Parameter settings

Although the proposed filter is improved based on the LCI detector, it reduces many unnecessary parameters than the original method. For the parameters σ _λ andσ _s in formula (1), and the parameters used to detect whether the noise pixel is in the detail or texture area, we have made fine adjustments based on the values given in the original literature, where the values of σ _λ andσ _s are respectively are 1.3 and 7.1. As for the classification parameter K, it is obvious that the more blocks the damaged image is divided into, the better the noise detection effect will be, but at the same time it will increase the time cost and complexity of the algorithm. We conducted experiments on pepper images with a noise level of 50 from 2 to 6 blocks to verify the impact of the number of blocks on noise detection. The experimental data are shown in Table 1. It can be found that when the value of the classification parameter is 4, various indicators Basically reach the optimal level.

Table 1 Noise detection situation when dividing different categories

K miss FALSE total psnr ssim 2 5314 15306 20620 27.83 0.86 3 5588 14378 19966 28.12 0.88 4 6130 12745 18875 28.66 0.92 5 6212 12443 18655 28.68 0.92 6 6266 12109 18375 28.71 0.93

2. Performance of noise detector

Since the detection accuracy of the noise detector has a great influence on the noise removal ability of the filter, a good noise detector should have fewer missed pixels, falsely detected pixels (MD and FD) and higher real detected noise. Pixel accuracy (true hit). Table 2 and Figure 5 show the detection effects of the proposed noise detector on test images with different noise levels. It can be seen from Table 2 that the effect will be better for some images with less texture and simpler pixel grayscale distribution, such as lena, pepper, prostate and brain images, whose MD and FD numbers are much less than other images. This is because noisy pixels in flat areas are easier to detect than noise pixels in edge areas. Although images such as baboon, Barbara, boat and bridge, which contain more details and textures, perform worse at low noise levels, it can be seen from Figure 5 that as the noise level increases, the detection rate of noise pixels in the image increases. Gradually improving, this is because there are fewer and fewer normal pixels in the image, and the grayscale distribution of pixels in the detection window is very different. If the central pixel is noise, then the difference between it and the pixels in its neighborhood will be more obvious. Therefore it is easier to detect as noise. Even when the noise level reaches 80%, the truth hit of almost all images is above 90%, which demonstrates the good stability and robustness of our proposed noise detector.

Table 2 RVIN. Detection results of this noise detector on different images from 30% to 80% RVIN

Generally, for grayscale images, it is not noticeable if the absolute difference between a pixel value and its neighboring pixel values is less than 8. In other words, when the gray value of the noise pixel differs within 8 from its original true value, it is difficult for the human eye or the noise detector to distinguish it, and their existence will not significantly reduce the image quality. , so for this part of noise pixels we can regard them as normal pixels. Based on this premise, we calculated the difference between the grayscale of the missed pixels of lena, pepper, barbara and gorilla with the noise level of 40% to 60% and their true values, as shown in Table 3, where D Represents the difference between the new value and the true value of the noise pixel. As can be seen from Table 3, the grayscale values of most of the missed noise pixels in these images are within 8 bits of their true values, which further verifies that the proposed noise detector has a high noise detection accuracy.

Table 3 The difference between the grayscale of missed pixels and the true value in different images with noise levels of 40% to 60%

In order to objectively evaluate the performance of the proposed noise detector, we compare it with several newly proposed and classic algorithms. The experimental results are shown in Table 4. It should be pointed out that for the FD and MD values of other noise detection algorithms, the best values mentioned in their literature were selected. It can be found from Table 4 that although some methods such as Luo,s have a low number of false positives, the number of missed pixels is very high, which will lead to more burrs in the image and affect the recovery performance of subsequent filters. The total number of the proposed noise detector is optimal under different noise levels. In fact, as the noise level increases, MD gradually reaches the optimum, which means that this method is very robust and the detector can still detect more noise pixels when the noise density becomes very high. Intuitively speaking, a good noise detector should be able to detect more noise pixels while reducing misjudgments as much as possible. Therefore, based on several evaluation indicators, we believe that the proposed noise detector has better performance than other methods. Better performance. In addition, it can also be found from the comparison results between LCI and the proposed filter that as the noise level increases, the effect gap between the proposed filter and LCI becomes more obvious, which shows that the improvements we proposed for LCI have obvious and substantial effects on detection performance. improvement.

Table 4 compares the detection results of RVIN contaminated Lena images under different noise levels.

3. Filter recovery performance on natural images

In order to verify the effectiveness and rationality of the proposed partition decision filter, we repair the lena image of 50% RVIN, as shown in Figure 6(a) and Figure 6(b). At the same time, a flat area A and a detailed area B with a size of 12×12 were selected from the image, and the corresponding pixel grayscale values before and after filtering are shown in Figure 6(c-f), where Figure 6(c) and Figure 6( d) Represents the pixel grayscale distribution of the flat area A and detail area B before filtering respectively. The pixels marked in gray are noise, and the values in brackets are their true values. Figure 6(e) and Figure 6(f) respectively represent Pixel grayscale distribution of flat area A and detailed area B after filtering. By zooming in on Figure 6(b), it can be found that there are no obvious burrs or residual noise clumps in the image. This is due to the fact that the proposed noise detector can well detect the noise in the damaged image. As shown in Table 3, although there are 6800 noise pixels that have not been detected, the gray value of half of the noise is very different from the original value, so it is not clearly observed visually. From the comparison in Figure 6(c-f), it can be seen that most of the noise pixels are detected whether in flat areas or detailed areas, and the restored pixel gray values are very close to the real values. Although for some normal pixels that are misjudged as noise, their new values are not much different from the original values after being filtered. It should be pointed out that the detection and filtering effects of detailed areas are significantly inferior to those of flat areas. For example, the 144 pixels in the flat area A after filtering all have an error of ±3 from the true value, while the error in the detailed area B is about 37 Each is above ±8. This is because the pixel grayscale distribution of edge and texture details is more complex. However, judging from the recovery effect in Figure 6(b), the proposed filter still retains and restores the texture and edges in the image better, without any obvious The image blur phenomenon shows the rationality and effectiveness of our regional filtering method.

In order to objectively evaluate the performance of the proposed filter, we choose PSNR and SSIM indicators to compare it with several mainstream filters, where PSNR is used to measure the similarities and differences between the original image and the reconstructed image, and SSIM is used to characterize the filter attention. The ability of details to preserve characteristics. It should be pointed out that for the PSNR and SSIM values of other filters, the best values mentioned in their literature were selected. From the PSNR comparison results in Table 5, we can find that except for the boat image, the proposed filter is slightly inferior to AEPWM, but in other images, especially at the noise level of 50 to 60, the performance of the proposed filter is more outstanding. At the same time, as the noise level increases, the peak signal-to-noise ratio value of the proposed filter attenuates more slowly than AEPWM and other filters. This is due to the fact that the noise detector we designed can detect more noise under high noise levels. There is a lot of noise, which shows that our method has good robustness. In the SSIM comparison results in Table 6, except for the bridge image which is slightly lower than ROR-NLM, the performance of the proposed filter on other images is significantly better than other filtering algorithms, which shows that the proposed filter can perform better Preserves edges and other detailed aspects in images.

Table 5 Comparison of peak signal-to-noise ratio recovery results for 40% to 60% RVIN images

Table 6 Comparison of the recovery effects of 40%～60%RVIN40%～60%RVIN images in SSIM

Similarly, several mainstream filtering algorithms were also selected to compare the effects of filters on visual output. As shown in Figure 7, it can be found that there are still many obvious noise images in ASMF, DWM, EAIF and SBF, while the restored image of SD-OOD has obvious blurring phenomenon, and the effects of ROR-NLM and AFWMF are relatively better. However, some edges and details are not well preserved, and the processing effects of BDND, EBDND and FRDFN are poor. In contrast, as shown in Figure 7(l), the results produced by the method used in this application have very good visual quality. Not only are there no obvious noise clumps and burrs in the image, but also by enlarging some detailed areas in the image It can be found that our filtering method preserves the lines and colors of the hull better than other methods, thanks to the high detection accuracy of the noise detector and the partitioned filtering design of the filter. It can be considered that for complex images with a noise density of 60%, our method can still detect and remove most of the noise pixels and retain most of the image details.

4. Restoration performance of filters on medical images

As shown in the filtering effect diagram of the prostate noise image shown in Figure 8 and the filtering effect diagram of the head noise image shown in Figure 9, it can be seen that the proposed denoising algorithm can restore different textures and resolutions under different RVIN intensities. Biomedical images. It can also be intuitively observed from the restored medical images that they have good texture and edge preservation capabilities, which helps ensure correct subsequent diagnosis and treatment.

The above description is only a preferred implementation of the RVIN detection and removal method based on pixel clustering disclosed in the present invention, and is not intended to limit the scope of protection of the embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of this specification shall be included in the protection scope of the embodiments of this specification.

Claims (9)

1. A method for RVIN detection and removal based on pixel clustering, comprising the steps of:

clustering and segmentation are carried out based on the gray distance similarity of the pixel points, and all pixels in the damaged image are classified into K classes;

calculating LCI values of pixels, determining an area where the pixels are located based on the LCI values, wherein the area comprises a flat area and a detail area, obtaining an optimal detection threshold value of each type of pixels through iterative solution, and judging whether the pixels are noise pixels or not according to the LCI values of the pixels and the optimal detection threshold value, wherein the LCI value calculating method comprises the following steps:

omega in equation (4) _x ⁰ Is a 5×5 neighborhood centered on pixel x, and y is a neighborhood Ω _x ⁰ Any one of the pixels, u _x And u _y The gray values (m, n) and (s, t) representing the pixels x and y are the coordinates, σ, of the pixels x and y, respectively _λ Sum sigma _s Parameters of a preset Gaussian kernel function are respectively set; θ (x, y) is the similarity between pixels x and y; ζx is a statistic for evaluating the likelihood that the center pixel x is a normal pixel; normalization by equations (6) and (7) limits the result to [0,1]]In the range of (2), the LCI value of the pixel x is finally obtained;

noise pixels for flat and detail regions are recovered with LCI weighted mean filters and edge direction filters, respectively, to recover pixels corrupted by random value impulse noise, as follows:

wherein I' _x A gray value representing the noise pixel x after filtering, y representing Ω determined to be non-noise in the noise detection stage _x ⁰ Pixels in (I) _y And LCI _y The gray value and LCI value of y are represented, respectively.

2. The method of pixel cluster-based RVIN detection and removal of claim 1, further comprising the steps of, prior to cluster segmentation:

the image is smoothed, which includes median filtering and gaussian filtering.

3. The method for detecting and removing RVIN based on pixel clustering according to claim 1, wherein the clustering segmentation is performed based on gray distance similarity of pixel points, and the clustering method adopts a K-means clustering method, and specifically comprises the following steps:

find K cluster centers μ _k (k=1, …, K) assigning all pixels in the corrupted image to the nearest cluster center such that the sum of squares of the one-dimensional distances of each pixel point from its corresponding cluster center, where one-dimensional distance refers to the gray difference of the two, introducing a binary variable r _nk E {0,1} to represent a pixel point x in the corrupted image _n For the assignment of cluster K (where n=1, …, N, k=1,..k), if pixel point x _n Belonging to the kth cluster, r _nk =1, otherwise 0, the following loss function can be defined:

from the above, it is known that the cluster center μ needs to be fixed randomly _k The initial value is used for obtaining the attribution value r of the pixel point minimizing the loss function J _nk Given pixel point x _n And cluster center mu _k Is the gray value of (1), the loss function J is r _nk Due to the linear function of x _n And x _n+1 Are mutually independent, for each pixel point x _n Only the point needs to be allocated to the nearest cluster center, i.e

Using r obtained in the formula (2) _nk Carrying out clustering center mu in formula (1) _k Given r _nk Is a value of mu, the loss function J is _k To make J vs. mu _k The derivative of (2) is 0, and can be obtained

Mu can be pushed out by the above method _k The value of (2) isμ _k The gray average value of the pixel points belonging to the class.

4. A method of pixel clustering based RVIN detection and removal as claimed in claim 3, wherein the clustering method employs mean shift clustering, density based clustering or maximum expected clustering of gaussian mixture models.

5. The method of claim 1, wherein the LCI values of the computed pixels are computed from the same class of pixels in the neighborhood of the pixel.

6. The method for RVIN detection and removal based on pixel clustering as claimed in claim 1, wherein said obtaining the optimal detection threshold for each type of pixel by iterative solution includes the steps of:

traversing the detection threshold from 0 to 1, calculating an objective function of the image denoising model, and when the objective function is minimum, the current detection threshold is the optimal detection threshold, wherein the objective function is as follows:

where y is any pixel of an image, i, j is the coordinate of y, V (y) is called TV norm, and is a regularization method for keeping image edge information as a target.

7. The method of pixel cluster based RVIN detection and removal of claim 1, wherein the LCI weighted mean filter has a filter window set to 5 x 5.

8. A method of RVIN detection and removal based on pixel clustering as claimed in claim 1, wherein the noisy pixels for detail areas employ an edge direction filter to recover pixels corrupted by random value impulse noise, comprising the steps of:

constructing a detection frame with noise pixels which are judged to be detail areas as centers;

the pixels which are distinguished as normal on the row, the column, the left diagonal and the right diagonal and are centered on the noise pixel in the detection frame are respectively put into a set D _h ,D _v ,D _l ,D _r In the collection;

respectively calculate D _h ,D _v ,D _l ,D _r The standard deviation of the elements in the set, and selecting the direction represented by the set with the smallest standard deviation as the boundary filtering direction;

the gray values of normal pixels in the boundary filtering direction are arranged in an ascending order or an inverse order, and the median in the election sequence is used as a new gray value of the central noise pixel.

9. The method of pixel cluster based RVIN detection and removal of claim 8, wherein the filtering window of the edge direction filter is set to 7 x 7.