CN102143303A

CN102143303A - Image denoising method in transmission line intelligent monitoring system

Info

Publication number: CN102143303A
Application number: CN 201110063780
Authority: CN
Inventors: 何冰; 刘新平; 沈超; 刘振海; 胡凌靖
Original assignee: SHANGHAI ELECTRIC POWER LIVE WORKING TECHNOLOGY DEVELOPMENT Co Ltd; Shanghai Municipal Electric Power Co
Current assignee: State Grid Corp of China SGCC; Shanghai Municipal Electric Power Co; Shanghai Jiulong Electric Power Group Co Ltd
Priority date: 2011-03-16
Filing date: 2011-03-16
Publication date: 2011-08-03

Abstract

The invention relates to an image denoising method in a transmission line intelligent monitoring system. The method is mainly used for filtering out noise in an image which is acquired by a video monitoring system and solving the problems of reduction of picture contrast and edge and the like under the conditions of dense fog, sand and dust and other severe weather. The image denoising method comprises the following steps of: smoothening the image by employing an anisotropy-based filter to estimate noise energy of each pixel point of the image so as to calculate a filtering residual function; taking the filtering residual function in a pixel neighborhood range as context features of the image; quantifying a context vector which corresponds to each pixel to 32 levels through a vector quantizer by employing a dynamic programming algorithm; and finally, constructing a filter with different parameters for all pixels on each level by employing a regression analysis method so as to enhance the image.

Description

Image denoising method in intelligent monitoring system of power transmission line

Technical Field

The invention relates to an image denoising method, and belongs to the technical field of image processing. In particular to an image denoising method in an intelligent monitoring system of a power transmission line.

Background

The video monitoring system has a lot of noises in the acquired video images due to insufficient light of the monitoring environment and severe weather and environment (such as sand storm), and the contrast and edges of the images are very blurred, while the video monitoring image noises are not a single noise generated by a noise model but a complex mixed noise, namely: random noise, quantization noise, and impulse noise. The most important difficulties encountered when applying the conventional image denoising method to the video image enhancement problem are: while the image noise is eliminated, some important image details or edge information is also weakened. This makes the design of the edge preserving filter a research focus for image processing. The traditional edge-preserving filter applies mathematical morphology operator, which can remove additive or multiplicative noise of image and enhance edge information, such as median filter. However, they still do not overcome the problem of some weak boundaries being weakened.

Due to the fact that noises with different properties exist in images obtained in severe weather and environment, the image quality of remote video monitoring images is often greatly reduced due to noise pollution and fuzzy contrast, the reliability and accuracy of video monitoring can be seriously affected, and the aims of timely alarming and early warning which are required by a monitoring system can not be achieved. Therefore, before sending the image to the monitoring center platform and the client, preprocessing such as noise reduction or enhancement is usually required to be performed on the video frame image. The usual enhancement is to apply a spatial filter to smooth the incoming noisy image within a local window of fixed size. However, such processing can seriously destroy the texture features and edge details of the image itself, and the contrast of the image can be reduced.

To solve this practical problem, many edge-preserving filters have been extensively studied and applied to software development of video surveillance images in recent years. For example, median filters, bilateral filters, anisotropic filters, etc. are widely used in various fields for image noise reduction, but most of them are effective only for gaussian noise or multiplicative noise, and are not ideal for solving the image enhancement problem with mixed noise.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the method is used for removing image noise and simultaneously preserving edge details and textures of an image.

In order to solve the technical problem, the invention adopts a context quantization technology in information theory to solve the estimation problem of the image noise model. Firstly, smoothing an image by using a filter based on gradient, and then calculating a filtering error energy function according to the obtained smoothed image, namely estimating high-frequency information on each pixel point, including edge information and noise around the pixel point. Then, a dynamic programming approach is applied to quantize the resulting error energy function to four different levels. Then, a set of quantized contexts is constructed after the resulting quantized error energies and texture features of the image. And finally, according to different quantization contexts, a regression analysis method is applied to construct filters with different parameters aiming at different context models, so that a self-adaptive filter is realized.

Specifically, the invention provides an image denoising method in an intelligent monitoring system of a power transmission line, wherein a given noise-free image is set as I, and a noise image to be processed is set as I_NoiseThe method comprises the following steps:

A. designing a gradient-based anisotropic filter, applying the filter to a given noise-free image I to obtain a filtered image

B. Based on the filtered image obtained in step ACalculating filtering residual error function g for each pixel point, and forming corresponding image context

C. According to the image context obtained in the step B

Constructing a 32-level vector quantizer

D. For each quantized context C_QSolving the filter coefficient b in the formula (1) by applying a regression analysis method_kAnd α, constructing a filter f (x | C) in a diamond-shaped window_Q)；

E. Applying the filter described in step A to a noisy image I_NoiseFiltering to obtain initial smooth image

F. Will I_NoiseAnd

subtracting to obtain residual g_NoiseAnd forming a corresponding image context

G. The context corresponding to each pixel is determinedBringing in

Obtaining the corresponding grade;

H. using filtersTo I_NoiseAnd filtering to obtain an output image.

A non-local edge retaining filter is calculated, important edge information is kept, and meanwhile noise of an image is removed to the maximum extent, so that the problems of denoising and enhancing in a video monitoring image are solved.

Preferably, the gradient-based anisotropic filter in step a is specifically designed as follows:

a1, calculating the gradient of a given noise-free image I

Difference in four directions

And

and corresponding filter coefficients c_Ni，j、c_Si，j、c_Wi，jAnd c_Ei，j：

A2, calculating the anisotropic filtering according to the parameters obtained above

As a result of (1):

a3, according to gradientDetermines the damping degree of filtering

{\hat{I}}_{i, j} = \frac{{\hat{I}}_{i, j} + I_{i, j}}{2}

else

{\hat{I}}_{i, j} = \frac{{3 \hat{I}}_{i, j} + I_{i, j}}{4} .

Further, the step B specifically includes:

b1 based on image I and filtered image

Computing a filtered error image

B2, using the high frequency information of the image in 3 x 3 where the current pixel point is as the image context to form the corresponding context vector

<math><mrow><mover><msub><mi>c</mi><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub><mo>&RightArrow;</mo></mover><mo>=</mo><mo>|</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>-</mo><mn>1</mn><mo>,</mo><mi>j</mi><mo>-</mo><mn>1</mn></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>-</mo><mn>1</mn><mo>,</mo><mi>j</mi></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>-</mo><mn>1</mn><mo>,</mo><mi>j</mi><mo>+</mo><mn>1</mn></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>,</mo><mi>j</mi><mo>-</mo><mn>1</mn></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>-</mo><mn>1</mn><mo>,</mo><mi>j</mi><mo>+</mo><mn>1</mn></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>+</mo><mn>1</mn><mo>,</mo><mi>j</mi><mo>-</mo><mn>1</mn></mrow></msub><mo>,</mo><msub><mi>g</mi><mrow><mi>i</mi><mo>+</mo><mn>1</mn><mo>,</mo><mi>j</mi></mrow></msub><msub><mi>g</mi><mrow><mi>i</mi><mo>+</mo><mn>1</mn><mo>,</mo><mi>j</mi><mo>+</mo><mn>1</mn></mrow></msub><mo>}</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>4</mn><mo>)</mo></mrow><mo>.</mo></mrow></math>

Still further, the step C specifically includes:

based on the obtained context vector

And filtering the error image g, applying a dynamic programming algorithm to minimize the values of the following conditional entropies:

<math><mrow><mo>-</mo><munderover><mi>Σ</mi><mrow><mi>d</mi><mo>=</mo><mn>1</mn></mrow><mi>m</mi></munderover><mi>p</mi><mrow><mo>(</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow><mi>Σp</mi><mrow><mo>(</mo><mi>g</mi><mo>|</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow><mi>log</mi><mi>p</mi><mrow><mo>(</mo><mi>g</mi><mo>|</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>5</mn><mo>)</mo></mrow></mrow></math>

wherein,

is the vector of the current pixel point in the image g relative to the context

Resulting conditional probability, i.e. solving a 32-level quantitizer

Vector context

Quantized into 32 bins of its defined space:

<math><mrow><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>=</mo><msubsup><mo>∪</mo><mrow><mi>d</mi><mo>=</mo><mn>1</mn></mrow><mn>32</mn></msubsup><mo>{</mo><msub><mi>C</mi><mi>Q</mi></msub><mo>|</mo><msub><mi>C</mi><mi>Q</mi></msub><mo>=</mo><mi>d</mi><mo>}</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>6</mn><mo>)</mo></mrow><mo>.</mo></mrow></math>

still further, the step D specifically includes:

d1, in the image I, for each pixel point y, taking a diamond neighborhood r (y) around it, as shown in fig. 2;

d2, obtaining a pixel point set satisfying Y ═ Y | C from the 32 quantized context sets C obtained in step C_Q(y)∈C，C_Q(y)＝C_QAnd has r (y) ═ r (y) | C_Q(y)∈C，C_Q(y)＝C_QAnd setting gray value y of pixel points and gray values x of other 12 pixel points in the rhombic neighborhood of all R (Y)_kThe following relationship is satisfied:

based on the obtained R (Y), we can analyze the coefficient b by regression_kAnd a, estimated coefficient b_kAnd α is also f (x | C)_Q) The filter coefficients of the filter.

Further, the step H specifically includes:

h1, in image I_NoiseIn the method, for each pixel point y, a diamond neighborhood r (y) is taken around the pixel point y, as shown in fig. 2;

h2, application filter f (x | C)_Q) To I_NoiseFiltering is carried out, namely:

the image denoising method based on the above comprises the following steps: a gradient-based anisotropic filter or predictor, a 32-level vector quantizerOne filter f (x | C)_Q). The functions and actions of the parts are described in the corresponding parts above.

The invention has the beneficial effects that: the invention applies the context quantization technology and the self-adaptive regression analysis method in the information theory to design a non-local edge-preserving filter, and removes the noise of the image to the maximum extent while keeping the important edge information, so as to solve the problems of denoising and enhancing in the video monitoring image. By designing the adaptive filter with an edge preserving, the edge detail and texture of the image are enhanced simultaneously while removing the image noise. Unlike the traditional noise estimation method, the noise estimation method by applying the context quantization technology in the information theory does not depend on the noise signal and the image signal, and the estimation method has strong robustness and is suitable for the unknown denoising problem of all noise models.

Drawings

FIG. 1 is a schematic flow chart of a two-dimensional image denoising method according to an embodiment of the present invention;

FIG. 2 is a diagram of a diamond neighborhood used in a spatial filter in an embodiment of the present invention.

Detailed Description

Hereinafter, the present invention will be described in more detail by way of examples with reference to the accompanying drawings. This example is merely a description of the best mode of carrying out the invention and does not limit the scope of the invention in any way.

Examples

FIG. 1 is a schematic flow chart of a two-dimensional image denoising method according to the present invention, which illustrates calculating a gradient image from an input noise-free training image sample 11, and designing an adaptive filter based on the obtained horizontal and vertical gradients, applying the filter to obtain a smoothed image 12; step 13 combines the original image I and the initial smoothed image

Obtaining high-frequency information of each pixel point, namely filtering residual errors, and forming context vectors by using residual error information in a neighborhood; step 14, quantizing the context vector into 32 levels according to a minimum conditional entropy criterion; step 15 applies regression analysis to estimate the filter coefficients under each set of context features from the quantized context features, constituting 32 filters corresponding to the contexts. The operations of

steps

22 and 23 are similar to those of

steps

12 and 13, and the difference is that initial filtering is performed on the input noise image, and corresponding context information is obtained; step 24, calculating the level of the current noise image context by using the quantizer calculated in step 14; step 31 performs filtering operation on the noise image according to the level of the context by using the 32 sets of filters obtained in step 15, and obtains an output noise reduction image 32. The method comprises the following specific steps:

1. calculating horizontal gradient and vertical gradient of a noise-free training image sample I, designing a gradient-adaptive anisotropic filter based on the horizontal gradient and the vertical gradient, and applying the filter to the image I to obtain a filtered image

2. Based on the resulting filtered image

Calculating a filtering residual error function g for each pixel point and generating a corresponding context vector

3. Based on the obtained context vector

Design 32-level quantizer of context vector by solving minimum conditional entropy

4. For each quantized context

Interval, applying regression analysis, constructing a filter in a diamond-shaped windowk＝1，2，Λ，32；

5. Noisy image to be processed I_NoiseObtaining a filtered image by applying the same filtering operation as the step 1

6. Will I_NoiseAndsubtracting to obtain residual g_NoiseAnd forming a corresponding image context

7. The context corresponding to each pixel is determined

Bringing in

Obtaining a corresponding grade;

8. using filters

To I_NoiseAnd filtering to obtain an output image.

The above process flow can be summarized as the establishment and invocation of the filter model. The establishment of the model is a core part and comprises three major parts, the estimation of image high-frequency information, the selection of an image context model and the design of a filter according to quantized context. The estimation of high frequency information of an image is usually a necessary process for designing an adaptive filter, and different noise models and estimation methods are different. The method is different from the most critical point of other technologies, the estimated high-frequency component quantity is not directly applied to the design of a filter, but is used for further describing and quantizing a context model of the characteristics around the image pixel points, and the method can be deduced according to the general source coding theory in the information theory: such a method has the advantage that a statistical model can be found that closely approximates the noise itself without knowing the noise model in the image.

The detailed steps of the invention are as follows:

calculating the gradient of the training image sample I

Difference in four directionsAndand corresponding filter coefficients c_Ni，j、c_Si，j、c_Wi，jAnd c_Ei，j：

Calculating anisotropic filtering according to the parametersAs a result of (1):

according to the gradient

Determines the damping degree of filtering

{\hat{I}}_{i, j} = \frac{{\hat{I}}_{i, j} + I_{i, j}}{2}

else

{\hat{I}}_{i, j} = \frac{{3 \hat{I}}_{i, j} + I_{i, j}}{4}

The anisotropic filter is a filter with different smoothing strengths designed according to different edge strengths and directions, and edge information in an image is well reserved while smoothing is performed. K. C is a parameter of the filter, K is responsible for controlling the degree of anisotropy, and may be a fixed value or may be dynamically determined from the histogram distribution of the gradient image, and C is used to further control the degree of filtering. Without loss of generality, K-64 and C-32 may be desirable.

Generally, the high frequency information of an image can be estimated by applying an adaptive filter, in other words, an adaptive filter

g may also be considered a somewhat noisy image, but also contains edge information and texture features. Using the high-frequency information of the image in 3 x 3 where the current pixel point is as the image context to form the corresponding context vector

Re-targeting image contextIs further estimated so that the noise model can be approximately described. According to Bayes's law and general source coding theory, the most approximate estimation of the image signal g is to solve the minimum conditional entropy value by an approximate model, namely, the minimum solving Kullback-Leibler distance:

<math><mrow><mo>-</mo><munderover><mi>Σ</mi><mrow><mi>d</mi><mo>=</mo><mn>1</mn></mrow><mi>m</mi></munderover><mi>p</mi><mrow><mo>(</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow><mi>Σp</mi><mrow><mo>(</mo><mi>g</mi><mo>|</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow><mi>log</mi><mi>p</mi><mrow><mo>(</mo><mi>g</mi><mo>|</mo><mi>Q</mi><mrow><mo>(</mo><mover><mi>c</mi><mo>&RightArrow;</mo></mover><mo>)</mo></mrow><mo>=</mo><mi>d</mi><mo>)</mo></mrow></mrow></math>

i.e. solving a vector quantizer

Will be provided with

Is divided into 32 intervals and arbitraryUniquely corresponding to one of these 32 intervals. In general, the problem of solving the minimum conditional entropy in a high-dimensional space is not convex, but we can apply a projective transformation to solve it

Projecting to a low-dimensional space to make the pair in the low-dimensional spaceThe resolution ratio of the method is the highest, the problem is converted into the problem of the minimum conditional entropy in a one-dimensional space, and the problem can be solved by a dynamic programming method.

Step 14 classifies each pixel into one of 32 models, so we will get 32 groups of data, each group of data contains the gray value of several current pixels y and the value of surrounding pixels in a diamond neighborhood, as shown in fig. 2, which illustrates the diamond neighborhood used by the spatial filter of the present invention, and also defines the pixels participating in the following regression analysis. For each set of data for diamond neighborhoods, the following assumptions may be made:

b_kis the estimated filter coefficients and alpha is the estimated approximate noise, x_kIs the gray value of the pixel point in the rhombus neighborhood of the current pixel point y. The filter coefficient b can be estimated by applying a regression analysis method based on the least square method_kAnd the value of the model noise a.

Steps 12 to 15 complete the filter model building, and in order to apply it to the actual noise reduction process, we perform the same initial filtering operation 22 as step 12 on the noise image 21 to be processed; subtracting the smooth image from the original image to obtain a filtering residual error, and forming a corresponding image context 23; substituting the context into the vector quantizer obtained in step 14

Obtaining its quantization level 24; and finally, filtering each pixel by using a corresponding filter to obtain an output image.

Claims

1. An image denoising method in an intelligent monitoring system of a power transmission line is provided, wherein a given noise-free image is set as I, and a noise image to be processed is set as I_NoiseThe method is characterized in that: the method comprises the following steps:

B. Based on the stepsThe filtered image obtained in step A

Calculating filtering residual error function g for each pixel point, and forming corresponding image context

C. According to the image context obtained in the step B

Constructing a 32-level vector quantizer

F. Will I_NoiseAnd

subtracting to obtain residual g_NoiseAnd forming a corresponding image context

G. The context corresponding to each pixel is determined

Bringing in

Obtaining the corresponding grade;

H. using filters

To I_NoiseAnd filtering to obtain an output image.

2. The image denoising method in the intelligent monitoring system of the power transmission line according to claim 1, characterized in that: the gradient-based anisotropic filter in step a is specifically designed as follows:

a1, calculating the gradient of a given noise-free image I

Difference in four directions

And

and corresponding filter coefficients CN_i，j、c_Si，j、c_Wi，jAnd c_Ei，j：

As a result of (1):

a3, according to gradient

Determines the damping degree of filtering

{\hat{I}}_{i, j} = \frac{{\hat{I}}_{i, j} + I_{i, j}}{2}

else

{\hat{I}}_{i, j} = \frac{{3 \hat{I}}_{i, j} + I_{i, j}}{4} .

3. The image denoising method in the intelligent monitoring system of the power transmission line according to claim 1, characterized in that: the step B specifically comprises the following steps:

b1 based on image I and filtered image

Computing a filtered error image

4. The image denoising method in the intelligent monitoring system of the power transmission line according to claim 1, characterized in that: the step C is specifically as follows:

based on the obtained context vectorAnd filtering the error image g, applying a dynamic programming algorithm to minimize the values of the following conditional entropies:

wherein,

is the vector of the current pixel point in the image g relative to the context

Resulting conditional probability, i.e. solving a 32-level quantitizer

Vector context

Quantized into 32 bins of its defined space:

5. the image denoising method in the intelligent monitoring system of the power transmission line according to claim 1, characterized in that: the step D is specifically as follows:

d1, in the image I, aiming at each pixel point y, a diamond neighborhood R (y) is taken around the pixel point y;

d2, obtaining the pixel point set satisfying Y ═ yC according to the 32 quantized context sets C obtained in step C_Q(y)∈C，C_Q(y)＝C_QAnd has r (y) ═ r (y) | C_Q(y)∈C，C_Q(y)＝C_QAnd setting gray value y of pixel points and gray values x of other 12 pixel points in the rhombic neighborhood of all R (Y)_kThe following relationship is satisfied:

6. The image denoising method in the intelligent monitoring system of the power transmission line according to claim 1, characterized in that: the step H is specifically as follows:

h1, in image I_NoiseIn the method, a rhombic neighborhood R (y) is taken around each pixel point y;

h2, application filter f (x | C)_Q) To I_NoiseFiltering is carried out, namely: