CN105678704B - A kind of non local intermediate value blind landing method for de-noising of view-based access control model perception - Google Patents

Abstract

The invention provides a non-local median blind noise reduction method based on visual perception, which includes the following steps: constructing an impulse noise blind detector based on the visual outlier measure of pixels in a digital image, and the visual outlier measure quantifies the impulse noise of different models The visual commonality is obtained by fusing the quantitative results of different visual features; the non-local information of the image is extracted, and the non-local median calculation model is constructed; the regularization parameters are calculated according to the visual outlier measure and non-local information, and the non-local median regularization item is established; Construct a non-local median denoising functional model to adaptively repair the noise pixels in the image. The blind noise reduction method of the present invention, according to the visual characteristics of the pulse noise, the self-similarity of the image itself, and outlier data mining, uniformly processes the pulse noise of different models and densities in the digital image, and can solve the problem of unknown pulse models and high noise in the actual noise reduction process. It is difficult to effectively repair the noisy pixels caused by density noise and image multimodal complexity.

Description

A Nonlocal Median Blind Denoising Method Based on Visual Perception

technical field

The present invention relates to the technical field of image processing, in particular to digital image noise reduction technology, in particular to a non-local median blind noise reduction method based on visual perception, which is suitable for blind noise reduction of unknown model impulse noise in digital images.

Background technique

Impulse noise is a common type of interference signal in digital images, which is caused by imperfections and errors in imaging systems, transmission media, and recording equipment during image acquisition, transmission, and storage. Impulse noise is usually divided into three types according to the distribution of brightness values, which are fixed value, random value, and mixed type of fixed value and random value. Suppressing impulse noise in digital images is the premise and foundation of image analysis, understanding and recognition, and it is also an important and difficult issue in this field. For the specific pulse model, domestic and foreign research institutes and researchers have carried out extensive research and obtained a large number of noise reduction methods. Generally speaking, they can be divided into two types: transform domain noise reduction and spatial domain noise reduction.

The idea of the transform domain noise reduction method is to transform the observed image, suppress the noise in the transform domain, and then obtain the final noise reduction result through inverse transformation. This type of method takes the distribution characteristics of the coefficients in the transform domain and the sparsity of the dictionary representation domain as a priori, and has strong multi-resolution and sparse representation capabilities, but the coefficients are complex to operate, strongly dependent on parameter settings and initial conditions, and usually have no Global solution, when repairing high-density noise images and complex images, it is easy to introduce false information and destroy the contrast, such as "ringing", "staircase", and "overlapping".

The spatial domain noise reduction method is to directly suppress the impulse noise in the spatial domain of the image. In contrast, the application of this method in the prior art is relatively mature, and the noise reduction result is closer to visual perception. The spatial domain removal methods of impulse noise can be roughly divided into two categories: linear and nonlinear. Mean filtering, median filtering and their improved algorithms are the most typical spatial filtering algorithms, but only using the mean value, median value and their simple deformation to restore noisy pixels, the assignment accuracy is low, which may lead to blurred noise reduction results or loss of detail information . Theory and experiments show that the regularized impulse noise removal method based on the energy functional model can effectively suppress the noise and preserve the details of the image more completely. Focusing on the design of the regularization model, the selection of regularization parameters, and the solution of the objective function, researchers at home and abroad have proposed l ₁ norm + edge-preserving regularization term, l ₁ norm + total variation term, l ₁ Excellent algorithms such as norm + partial differential constraint, l ₁ norm + l _p norm constraint term. In general, most of these methods have ideal noise reduction performance, can suppress pulses and effectively preserve image details, but the premise is that the prior constraints are accurate and reliable, and the regularization parameters are selected reasonably. When selecting regularization parameters, most of the current algorithms adopt a unified definition in advance and then optimize through a large number of experiments. However, defining consistent parameter values for noise pixels with different characteristics in the image makes the fidelity and smoothness unbalanced in complex areas and high-density noise areas of the image, and the accuracy of complex image and high-density noise image restoration is reduced.

In summary, most of the existing methods can obtain better noise reduction results when the pulse model and noise density are known, and the image to be repaired is relatively simple. However, considering that in the actual noise reduction process, the specific model and density of the impulse noise in the image and the complexity of the image to be repaired are rarely known in advance, so it involves the blind detection of the unknown model impulse noise in the image and the detection of pixels in different feature regions. It is difficult for existing noise reduction methods to effectively deal with problems such as adaptive detection, repair, and effective removal of high-density impulse noise.

Contents of the invention

Aiming at the defects or deficiencies in the prior art, the present invention aims to propose a non-local median blind denoising method based on visual perception, which can effectively suppress the impulsive noise model, noise density and image complexity. Noise, and completely protect the details of the image.

Another object of the present invention is to provide a visual perception-based blind noise reduction device for impulse noise, and a computer system for realizing the aforementioned visual perception-based non-local median blind noise reduction.

The above objects of the invention are achieved by the technical features of the independent claims, which the dependent claims develop in an alternative or advantageous manner.

In order to achieve the above object, the first aspect of the present invention proposes a non-local median blind noise reduction method based on visual perception, including the following steps:

Step 1. Constructing an impulse noise blind detector based on the visual outlier measure of pixels in the digital image. The visual outlier measure is obtained by quantifying the visual commonality of impulse noise of different models and fusing the quantitative results of different visual features;

Step 2, extracting the non-local information of the image, and constructing a non-local median calculation model;

Step 3. Calculate the regularization parameters based on the visual outlier measure and non-local information, and establish a non-local median regularization term;

Step 4. Establish a non-local median denoising functional model based on steps 2 and 3, and adaptively repair noise pixels in the image.

According to the present disclosure, another aspect of the present invention also proposes a blind noise reduction device based on visual perception of impulse noise, including:

A first module for constructing a blind impulse noise detector based on a visual outlier measure of pixels in a digital image, wherein the visual outlier measure is obtained by quantifying the visual commonality of different model impulse noises and fusing the quantitative results of different visual features;

It is used to extract the non-local information of the image and construct the second module of the non-local median calculation model;

A third module for calculating regularization parameters based on visual outlier measures and non-local information, and establishing a non-local median regularization term;

It is used to construct a non-local median noise reduction functional model based on the non-local median calculation model constructed by the aforementioned second module and the non-local median regularization term established by the third module. The non-local median noise reduction functional The model is configured to adaptively inpaint noisy pixels in images.

According to the improvement of the present invention, the third aspect of the present invention also proposes a computer system for implementing non-local median blind noise reduction based on visual perception, the computer system comprising:

memory;

one or more processors;

one or more modules stored in the memory and configured to be executed by the one or more processors, the one or more modules including a module for performing :

A first module for constructing a blind impulse noise detector based on a visual outlier measure of pixels in a digital image, wherein the visual outlier measure is obtained by quantifying the visual commonality of different model impulse noises and fusing the quantitative results of different visual features;

It is used to extract the non-local information of the image and construct the second module of the non-local median calculation model;

A third module for calculating regularization parameters based on visual outlier measures and non-local information, and establishing a non-local median regularization term;

It is used to construct a non-local median noise reduction functional model based on the non-local median calculation model constructed by the aforementioned second module and the non-local median regularization term established by the third module. The non-local median noise reduction functional The model is configured to adaptively inpaint noisy pixels in images.

Compared with the prior art, the blind noise reduction scheme proposed by the present invention has significant beneficial effects:

1. Quantify and integrate the outlier characteristics of different models of impulse noise from a visual perspective, propose a pixel visual outlier measure, and construct an impulse noise blind detector to uniformly distinguish impulse noise of different models, and realize blind detection of impulse noise of unknown models;

2. Designed a non-local median impulsive noise removal method, adaptive regularization parameters, combined with non-local median to construct a noise reduction functional model, and increased the prior constraints of the objective function, thereby improving the repair accuracy of noisy pixels;

3. In the case of unknown impulse noise model, noise density and image complexity, it can effectively suppress noise and completely protect the details of the image.

It should be understood that all combinations of the foregoing concepts, as well as additional concepts described in more detail below, may be considered part of the inventive subject matter of the present disclosure, provided such concepts are not mutually inconsistent. Additionally, all combinations of claimed subject matter are considered a part of the inventive subject matter of this disclosure.

The foregoing and other aspects, embodiments and features of the present teachings can be more fully understood from the following description when taken in conjunction with the accompanying drawings. Other additional aspects of the invention, such as the features and/or advantages of the exemplary embodiments, will be apparent from the description below, or learned by practice of specific embodiments in accordance with the teachings of the invention.

Description of drawings

The figures are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures may be represented by a like reference numeral. For purposes of clarity, not every component may be labeled in every drawing. Embodiments of the various aspects of the invention will now be described by way of example with reference to the accompanying drawings, in which:

Fig. 1 is a flow chart of a non-local median blind noise reduction method based on visual perception according to some embodiments of the present invention.

Figures 2a-2d are schematic diagrams of images of two types of pulse interference (both noise densities are 30%).

Figures 3a-3c are schematic diagrams of X-ray images interfered with by 50% random value noise and the results of noise reduction processing, respectively.

Figures 4a-4c are schematic diagrams of tongue coating images subjected to 70% fixed-value noise and the results of noise reduction processing, respectively.

Detailed ways

In order to better understand the technical content of the present invention, specific embodiments are given together with the attached drawings for description as follows.

Aspects of the invention are described in this disclosure with reference to the accompanying drawings, which show a number of illustrated embodiments. Embodiments of the present disclosure are not necessarily intended to include all aspects of the invention. It should be appreciated that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of numerous ways, since the concepts and embodiments disclosed herein are not limited to any implementation. In addition, some aspects of the present disclosure may be used alone or in any suitable combination with other aspects of the present disclosure.

According to the embodiment of the present invention, in general, the non-local median blind noise reduction method based on visual perception proposed by the present invention, according to the visual characteristics of impulse noise, self-similarity of the image itself and outlier data mining, uniformly processes digital images The impulsive noise of different models and densities in the medium aims to solve the problem that the noise pixel is difficult to be effectively repaired due to the unknown impulsive model, high-density noise, and multi-modal complexity of the image in the actual noise reduction process.

The whole process of blind noise reduction roughly includes two stages, which are: 1) Construct a blind impulsive noise detector based on the visual outlier measure of pixels in the digital image, and blindly discriminate the impulsive noise of the unknown model in the digital image; 2) In the noise In the pixel assignment stage, the target functional model is established based on the non-local median noise reduction algorithm, and the noise pixels in the image are adaptively repaired.

The aforementioned visual outlier measure calculates the visual outlier measure of each pixel by quantifying the visual commonality of different models of impulse noise and combining the quantification results of different visual features to provide a measurement basis for the detection of noise pixels.

As described in the following specific embodiments, the visual commonality of noise impulse noise of different models is mainly reflected in three aspects: isolated spatial distribution, poor connectivity, and abnormal brightness. The blind noise reduction method of the present invention aims to use these visual commonalities to quantify, fuse the quantified results, and calculate the visual outlier measure of the pixels in the digital image.

On the basis of fusion, we build a blind detector to blindly distinguish the unknown model impulse noise in the digital image. Then, the noise in the digital image is adaptively repaired by using the target functional model based on the non-local median noise reduction algorithm.

The implementation of the blind noise reduction method of the foregoing embodiment will be described in more detail below in conjunction with the accompanying drawings.

As shown in FIG. 1, according to an embodiment of the present invention, a non-local median blind noise reduction method based on visual perception includes the following steps:

Step 1. Constructing an impulse noise blind detector based on the visual outlier measure of pixels in the digital image. The visual outlier measure is obtained by quantifying the visual commonality of impulse noise of different models and fusing the quantitative results of different visual features;

Step 2, extracting the non-local information of the image, and constructing a non-local median calculation model;

Step 3. Calculate the regularization parameters based on the visual outlier measure and non-local information, and establish a non-local median regularization term;

Step 4. Establish a non-local median denoising functional model based on steps 2 and 3, and adaptively repair noise pixels in the image.

As an optional example, the implementation of the aforementioned step 1 includes two processes of quantifying the visual commonality of impulse noise, fusing quantization results, and building a blind detector, which will be described in detail below.

First, combining the visual commonality characteristics of noise impulse noise of different models, our processing of visual commonality quantification of impulse noise includes: spatial outlier quantization of pixels and luminance outlier quantization of pixels.

a. Spatial outlier quantification of pixels: According to the data characteristics of local connectivity of digital images, using spatial outlier measure and adopting connectivity-based anomaly mining algorithm to calculate the spatial measure (IM:isolationmeasurement) IM(i) of any pixel i ;

b. Luminance outlier quantification of pixels: based on the Weber-Fechner law, the minimum perceivable difference in luminance in the local area of the target pixel is studied, and combined with the local spatial outlier measurement, the luminance outlier measure of pixel i relative to its background is quantified (LTM : luminance transition measurement) LTM(i).

Combining the results of the aforementioned spatial outlier quantization of pixels and pixel brightness outlier quantification, we can obtain a visual outlier measure (VPOM: visual perception outlier measurement) VPOM(i) for each pixel, and on this basis construct a pixel-based Impulse-noise blind detector for visual outlier metrics.

Since the calculation model of the non-local mean denoising algorithm comes from solving the minimum value of the function with P as the independent variable.

min _P (Σ _j∈P(i) ω _i,j |PP _j | ² )

In the formula, i represents the target pixel, j represents the non-local pixel, P _j is the image block centered on j, P(i) is the self-similar pixel search window of pixel i, ω _i,j is the similarity between pixel i and j .

Considering the nonlinear characteristics of impulse noise, converting the mean solution in the non-local mean algorithm to the median solution can improve the assignment accuracy of noise pixels, and find the weighted median value of all self-similar pixels of the target pixel i—non-local median algorithm , solve P for the minimum value of the independent variable function to obtain the median value.

Therefore, in the aforementioned step 2, we construct the non-local median calculation model in the following way:

min _P (Σ _j∈P(i) ω _i,j |PP _j | ² )

As mentioned above, i represents the target pixel, j represents the non-local pixel, P _j is the image block centered at j, P(i) is the self-similar pixel search window of pixel i, ω _i,j is the similarity between pixel i and j Spend.

Therefore, the P obtained by solving is the minimum value of the independent variable function, that is, the non-local median value is obtained.

At the same time, since noise pixels contribute little to the repair of i, in some embodiments, the weight ω _i,j is optimized by combining the outlier measure of the pixel and the fuzzy membership degree of the pixel. In order to quickly solve the corresponding weighted mean value, as a preferred example, the method of setting a threshold to limit the number of self-similar pixels is used to solve the minimum value, so as to reduce the computational complexity and improve the convergence speed of the computational model.

Of course, in some other embodiments, other existing known methods can also be used to solve the problem, which will not be repeated here.

We use the energy functional model as the framework, use the non-local median to construct the regular term, and establish the noise reduction functional model, which will be described in more detail in the following content.

In step 3, we analyze the local information of the area where the pixel is located based on the visual outlier degree of the pixel and the detection result of the noise pixel (the number of times T(i) that the pixel i is judged to be a noise pixel during the iterative process). Adaptively determine the regularization parameter:

λ(i)=λ ₀ f ₁ (VPOM(i))f ₂ (T(i))

_λ0 is the initial regularization parameter, and _f1 and _f2 are weight functions.

In step 4, we combine the local regularization term, the non-local median calculation model and the adaptive regularization parameter to construct the non-local median regularization term.

In the following, as shown in FIG. 1 , an exemplary description will be given of the process of performing blind noise reduction processing on digital images using the non-local median blind noise reduction method based on visual perception in the foregoing embodiments.

Step 1. Input an observation image u disturbed by impulse noise

The image u input in this step is disturbed by impulse noise, but the specific impulse model is unknown, it may be a fixed value impulse model or a random value impulse model, and of course it may even be a mixed model of the two. As in the example of Figures 2a-2d, the noise density of the pictures in the figures is 30%.

Step 2. Calculate the spatial and brightness outlier measures of any pixel i in the observed image u, and fuse the calculation results to obtain the visual outlier measures of each pixel in the image

a. Compute the spatial outlier measure for any pixel i in image u

According to the human eye’s visual perception of brightness in the 9×9 area centered on pixel i, the variable threshold LUT(u _l ) of the following formula is used, where l=i+k,k∈[-4,4], to calculate For the connected pixel chain of the pixel, the connectivity parameter C of the pixel is defined by the largest connected pixel chain, that is, the pixel number of the pixel chain containing the largest number of pixels.

In the formula, u _l represents the brightness of the current pixel l in the pixel chain, and LUT(l) is a variable threshold with the brightness of pixel l as the background.

Calculate the 10 pixels with the smallest brightness difference from pixel i in the 5×5 window centered on pixel l, find out the connectivity parameters of these pixels, take the median value C1, and then take the connectivity of all pixels in the entire 5×5 window The median value C2 of the sex parameter, take the ratio of the two as the connectivity measure IM(i) of the pixel i.

b. Compute the brightness outlier measure for any pixel i in image u

According to the spatial outlier measure, in a 5×5 window centered on pixel i, the α clipping mean value of the image block is calculated as the background brightness of the local area according to the following formula:

In the formula, u _α is the mean value of α clipping, n is the number of pixels in the image block, u _k is the kth value after arranging n pixels from small to large, here α=18.

In a 5×5 window centered on pixel i, calculate the 10 pixels with the smallest brightness difference from pixel i, and calculate the brightness difference S _t ,t∈[1,10] between these pixels and pixel i, according to Fechner The law calculates the local visual brightness difference of pixels, as follows:

c. Combining the luminance outlier measure and the spatial outlier measure, calculate the visual outlier measure VPOM(i) of any pixel i in the image u, and this value is the basis for judging whether the pixel i belongs to a noise pixel.

VPOM(i)=β·IM(i)+γ·LTM(i)

In the above formula, β and γ are the fusion coefficients of the brightness outlier measure and the space outlier measure, where β=γ=0.5.

Step 3. Using the visual outlier measure of each pixel in the image u, construct a blind detector with the following formula, and detect noise pixels in the image by threshold T ^k :

T ^k ＝T ^k-1 ·0.9, k＝1,2,3,...K _max

In the above formula, k is the number of iterations (the noise reduction process of the present invention adopts iterative processing), and K _max is the maximum number of iterations.

Step 4. Extract the non-local information of any pixel i in the image u

Select a 21×21 image block with pixel i as the center, and use the following kernel function to calculate the self-similar pixel weight ω _{i,j of} pixel i in this pixel block:

In the above formula, u _i and u _j are the pixel values of i and j respectively, and λ=16.

Step 5, calculate the regularization parameter of any pixel i in the image u

Based on the visual outlier of the pixel and the detection result of the noisy pixel (the number of times T(i) that the pixel i is judged as a noisy pixel in the iterative process), the regularization parameter is adaptively determined for the pixel.

λ(i)=λ ₀ f ₁ (VPOM(i))f ₂ (T(i))

λ ₀ is the initial regularization parameter, here λ ₀ =0.01, f ₁ and f ₂ are weight functions, where,

In the above formula, VPOM(i+k) is the visual outlier measure value of the pixel in the 3×3 area centered on the pixel i.

In the above formula, K _max represents the maximum number of iterations in the noise reduction process.

Step 6, establish the target noise reduction functional model F _r : R ^M×N → R, and estimate and repair the pixels to be repaired in the image u (noise pixels detected in step 3)

In the above formula, V≡{1,2,...,M}×{1,2,...,N}, which represents an image with a size of M×N, r represents the noise reduction repair image, where i represents the target pixel, j is the self-similar pixel of pixel i, Q is the 49 most similar pixels to i among the self-similar pixels, r _i represents the repair result of pixel i, r _j represents the repair result of pixel j, and the value of λ is 16.

The noise pixel in the image is estimated and repaired by finding the minimum value of F _r (u). The noise reduction in this embodiment is an iterative algorithm, which gradually detects and repairs the impulse noise in the image through iteration, and finally restores the image u .

As the blind noise reduction method described in the foregoing embodiments, the noise reduction effect thereof will be further described below in conjunction with some examples of using the method to perform noise reduction processing on digital images.

1) The experimental conditions are windows8, CPU Inter(R)Core(TM)i5, 2.5GHz, and the software platform is Matlab7.9.1.

The first data selected for the simulation is the X-ray image polluted by 50% random value noise, as shown in Figure 3a, and the second data is the tongue coating image 4a polluted by 70% fixed value noise. The third data is Lena image, Baboon image, Goldhill image, Boat image, Pepper image interfered by 30% mixed noise.

2) Experimental content and results

Under the above experimental conditions, the traditional non-local mean method (NLM method) and the method of the foregoing embodiments were used to process the first and second experimental data respectively. The noise reduction results obtained by the NLM method are shown in Figure 3b and Figure 4b; the noise reduction results obtained by the present invention are shown in Figure 3c and Figure 4c.

Comparing Fig. 3b, Fig. 3c and Fig. 4b, Fig. 4c, it can be seen that the NLM method of the prior art has a serious loss of image detail loss, and a certain degree of damage appears on the edge of the image. The method of the aforementioned embodiment of the present invention can not only effectively It can remove the noise while maintaining the details of the image, which is obviously better than the NLM method of the prior art in terms of visual effect.

Under the above experimental conditions, use the NLM method and the method of the present invention to denoise the third data, and calculate the overall peak signal-to-noise ratio PSNR and mean absolute error MAE of the denoising result image, and the results are shown in Table 1.

Table 1—PSNR value and MAE value of image inpainting results interfered with by 30% impulse noise

It can be seen from Table 1 that the noise reduction effect of the method of the foregoing embodiments of the present invention is significantly better than that of the NLM algorithm in the mixed model impulse noise, and the PSNR value reflecting the image quality is relatively high, and the MAE value reflecting the loss of image detail is small.

In summary, compared with the traditional NLM method, the method of the present invention has obvious advantages in removing various types of impulse noise, and the noise reduction performance is better. The PSNR of the resulting image is significantly improved, the detail protection is more complete, and the unknown impulse Blind detection of model noise and adaptive inpainting of noisy pixels in high-density noisy images and complex images improve the accuracy of noisy pixel inpainting.

According to the present disclosure, the present invention also relates to a blind noise reduction device based on visual perception of impulse noise, comprising:

A first module for constructing a blind impulse noise detector based on a visual outlier measure of pixels in a digital image, wherein the visual outlier measure is obtained by quantifying the visual commonality of different model impulse noises and fusing the quantitative results of different visual features;

It is used to extract the non-local information of the image and construct the second module of the non-local median calculation model;

A third module for calculating regularization parameters based on visual outlier measures and non-local information, and establishing a non-local median regularization term;

It is used to construct a non-local median noise reduction functional model based on the non-local median calculation model constructed by the aforementioned second module and the non-local median regularization term established by the third module. The non-local median noise reduction functional The model is configured to adaptively inpaint noisy pixels in images.

It should be understood that the functions, functions and effects of the first module, the second module, the third module and the fourth module proposed in this embodiment have been described in the above description of the non-local median blind noise reduction method based on visual perception For the sake of illustration, its implementation has been exemplified in the aforementioned embodiments of the blind noise reduction method, and will not be repeated here.

According to the foregoing embodiments of the present invention, such as a non-local median blind noise reduction method based on visual perception and a non-local median blind noise reduction device based on visual perception, the present invention also proposes a method for realizing non-local median blind noise reduction based on visual perception A computer system for median blind noise reduction, the computer system comprising:

memory;

one or more processors;

one or more modules stored in the memory and configured to be executed by the one or more processors, the one or more modules including a module for performing :

A first module for constructing a blind impulse noise detector based on a visual outlier measure of pixels in a digital image, wherein the visual outlier measure is obtained by quantifying the visual commonality of different model impulse noises and fusing the quantitative results of different visual features;

It is used to extract the non-local information of the image and construct the second module of the non-local median calculation model;

A third module for calculating regularization parameters based on visual outlier measures and non-local information, and establishing a non-local median regularization term;

It is used to construct a non-local median noise reduction functional model based on the non-local median calculation model constructed by the aforementioned second module and the non-local median regularization term established by the third module. The non-local median noise reduction functional The model is configured to adaptively inpaint noisy pixels in images.

It should be understood that the aforementioned memory is used to store programs and data for execution by the processor. These memories, for example, may be a memory with a magnetic disk as a storage medium, or a memory based on a flash memory chip, and the like.

Obviously, in the computer system of this embodiment, these stored modules can be executed by one or more processors to realize the blind noise reduction process described in the aforementioned non-local median blind noise reduction method based on visual perception, so as to achieve the recognition of unknown Blind detection of impulse model noise and adaptive restoration of noise pixels in high-density noise images and complex images, improving the accuracy of noise pixel restoration.

Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the claims.

Claims (1)

1. A non-local median blind noise reduction method based on visual perception is characterized by comprising the following steps:

step 1, constructing an impulse noise blind detector based on visual outlier measurement of pixels in a digital image, wherein the visual outlier measurement is obtained by quantizing the visual commonality of impulse noise of different models and fusing different visual feature quantization results;

step 2, extracting non-local information of the image and constructing a non-local median calculation model;

step 3, calculating regularization parameters according to the visual outlier measure and the non-local information, and establishing a non-local median regularization item;

step 4, establishing a non-local median noise reduction functional model according to the steps 2 and 3, and adaptively repairing noise pixels in the image;

in step 1, the specific implementation of constructing the impulse noise blind detector based on the visual outlier measure of the pixels in the digital image includes:

1-1) calculating the spatial outlier measure of any pixel i in an image u interfered by impulse noise;

1-2) calculating the brightness outlier measure of any pixel i in the image u;

1-3) fusing the brightness outlier measure and the space outlier measure, and calculating the visual outlier measure VPOM (i) of any pixel i in the image u, wherein the value is a basis for judging whether the pixel i belongs to a noise pixel; and

1-4) Using the visual outlier measure VPOM (i) of each pixel in the image u, a blind detector is constructed with the following formula, passing a threshold T^kDetecting noisy pixels in an image:

T^k＝T^k-1·0.9,k＝1,2,3,…K_max

in the above formula, K is the number of iterations, K_maxIs the maximum number of iterations;

in step 1-1), the spatial outlier measure of any pixel i is calculated as follows:

based on the visual perception of brightness by the human eye in the 9 × 9 neighborhood centered on pixel i, a variable threshold LUT (u) of the formula_l) Where l ═ i + k, k ∈ [ -4,4]Calculating a connected pixel chain of the pixel, and defining a connectivity parameter C of the pixel by using the largest connected pixel chain, namely the number of pixels of the pixel chain containing the largest number of pixels:

in the formula u_lRepresenting the luminance of the current pixel l in the pixel chain, LUT (u)_l) A variable threshold against which the brightness of pixel l is background;

calculating 10 pixels with the minimum brightness difference with the pixel i in a window which takes the pixel l as the center and is 5 multiplied by 5, finding out the connectivity parameter of the pixels, taking the median value C1, then taking the median value C2 of the connectivity parameters of all the pixels in the whole 5 multiplied by 5 window, taking the ratio of the two as the connectivity measure IM (i) of the pixel i, wherein the connectivity measure IM (i) is the spatial outlier measure of the pixel i;

in addition, the luminance outlier metric in the step 1-2) is calculated as follows:

according to the spatial outlier measure, in an image block 5 × 5 centered on a pixel i, an α -clipped mean of the image block is calculated as the background luminance of the local area according to the following formula:

in the formula u_αis α clipping mean, n is the number of pixels in the image block, u_kis the kth value after arranging n pixels from small to large, α being 18;

in a window of 5 × 5 with the pixel i as the center, 10 pixels having the smallest luminance difference with the pixel i are calculated, and the luminance difference S between these pixels and the pixel i is calculated_t,t∈[1,10]The local visual luminance difference of the pixel is calculated according to fisher's law as follows:

the local visual brightness difference LTM (i) obtained by calculation is the brightness outlier measure of the pixel i;

and in step 1-3), fusing the luminance outlier measure and the spatial outlier measure by using the following formula to calculate the visual outlier measure VPOM (i) of any pixel i in the image u:

VPOM(i)＝β·IM(i)+γ·LTM(i)

in the above formula, β and γ are fusion coefficients of luminance outlier measure and spatial outlier measure, and β ═ γ ═ 0.5 is taken;

the specific implementation of the step 2 comprises:

selecting a 21 × 21 image block centered on a pixel i, and calculating a self-similarity pixel weight ω of the pixel i in the image block by using the following kernel function_i,j：

In the above formula, u_iAnd u_jPixel values of i and j, respectively, λ ═ 16;

the specific implementation of the step 3 comprises:

and (3) adaptively determining a regularization parameter lambda (i) for the pixel according to the visual outlier of the pixel and the detection result of the noise pixel:

λ(i)＝λ₀f₁(VPOM(i))f₂(T(i))

λ₀is an initial regularization parameter, here taken as λ₀＝0.01，f₁And f₂Is a weight function, and T (i) is the detection result of the noise pixel, i.e. the number of times the pixel i is judged as the noise pixel in the iteration process;

wherein,

in the above equation, VPOM (i + k) is the visual outlier measure of the pixels in the 3 × 3 neighborhood centered on pixel i;

in the above formula K_maxRepresenting the maximum number of iterations in the noise reduction process.