CN111179286A - FCM rotary kiln flame image segmentation method based on energy noise detection - Google Patents

Abstract

The accurate judgment of the combustion state in the rotary kiln is one of the important steps in the production of the rotary kiln, and the accurate identification of the combustion condition of the flame of the rotary kiln can not only increase the production efficiency of the rotary kiln, save the cost and improve the production quality of products, but also ensure the personal safety of workers and reduce the discharge amount of pollutants in the production process. With the continuous and deep research on computer software technology, the problem of identifying the flame image of the rotary kiln by using the computer technology is one of the main modes of the current industrial production. The method is used for researching the identification of the flame image of the rotary kiln, analyzing the application of image segmentation in the confluent flame image of the rotary kiln and segmenting the flame image of the rotary kiln by using a common algorithm. The method aims at improving poor flame image segmentation effect of the traditional fuzzy clustering algorithm (FCM), the improved algorithm is applied to the flame image, and effectiveness of the improved algorithm and accuracy of flame image segmentation are obtained through experimental analysis and comparison.

Description

FCM rotary kiln flame image segmentation method based on energy noise detection

Technical Field

The identification technology of the rotary kiln has many problems, and how to improve the segmentation identification of the shot flame image by using an intelligent optimization algorithm has very important practical significance.

Background

With the rapid development of the computer industry, image segmentation has become one of the main methods for identifying the flame status during the operation of the rotary kiln. And selecting and extracting characteristic values of the flame image in the kiln by utilizing image segmentation, and then identifying the target flame by utilizing the segmented characteristic values so as to obtain the flame state in the rotary kiln. In recent years, many scholars at home and abroad deeply dig out the application of the algorithm to pattern recognition and recognize and divide the rotary kiln flame image by using image division. The traditional image segmentation method is low in speed and complex in calculation, cannot monitor the flame combustion state in the kiln in real time, and can only segment the acquired flame image at a later stage. However, due to the influence of the combustion environment in the rotary kiln, the state of the flame is easily interfered by dust in the kiln when shooting, and the obtained image is seriously polluted, so that the segmentation is difficult.

Aiming at the defects of the traditional FCM, scholars at home and abroad carry out various improvement researches on the FCM and make a series of progresses. ChenS.H and the like use spatial neighborhood information to constrain an objective function, and a series of algorithms such as FCM-S and the like are provided to improve the segmentation precision, but the algorithm efficiency is low. The Kangjiayin filters the image based on the self-adaptive weighted mean value and corrects the objective function, and provides an improved FCM considering pixel space information, but the algorithm ignores detail information in the image. By analyzing the sample spatial distribution characteristics and the relation thereof, the Xiaochun and the like provide an improved algorithm which gives consideration to image detail information by means of FCM based on spatial correlation, but the algorithm has a large calculation amount. Hocrown English and other algorithms combining space constrained fast FCM and Markov random fields to realize side scan sonar image segmentation, but image detail information is poor to keep. Guo F F and the like propose an adaptive FCM segmentation method based on local noise detection on the basis of documents, and the algorithm distinguishes noise by calculating the variance of neighborhood gray values, but introduces more parameters and easily loses the detailed information of images.

Compared with the image segmentation of other algorithms, the FCM (ENFCM) algorithm based on energy noise detection has certain advantages, has the advantages that an energy curve is introduced to achieve adaptivity of parameters, the target function is updated by using the distance between samples to update the membership degree and the clustering center to improve the spatial correlation of pixels, and the segmentation accuracy and the original information of the image are well reserved.

Disclosure of Invention

The invention provides an FCM algorithm based on energy noise detection on the basis of NDFCM and is used for realizing initial segmentation of a rotary kiln flame image. The NDFCM algorithm obtains a good anti-noise effect by combining with a spatial neighborhood constraint condition, but the method needs to consider and select a proper window to prevent the loss of image detail information when setting parameters. In order to overcome errors caused by a window in the image segmentation process, the FCM algorithm based on energy noise detection introduces data information of an energy curve compressed image by combining spatial information so as to realize accurate and efficient segmentation of a rotary kiln flame image.

The optimization algorithm comprises the following steps:

the method comprises the following steps: let I ═ l_ijWhere 1. ltoreq. i.ltoreq.m, 1. ltoreq. j.ltoreq.n is an image of size mxn, where l_ijIs the gray value of the image I at pixel location (I, j). L is the maximum gray value of the image I. Defining the d-order neighborhood system of (i, j) as

A second-order neighborhood system is utilized, i.e., (u, v) ∈ { (+ -1, 0), (0, + -1), (1, + -1), (-1, + -1) }. Two-dimensional matrix B generated from I image_l＝{b_ij,1≤i≤m,1≤j≤n}，C＝{c_ijI is more than or equal to 1 and less than or equal to m, j is more than or equal to 1 and less than or equal to n, where c_ijWhen the energy curve is generated, different positions are adopted according to a d-order neighborhood system, and an energy curve function associated with each gray value in the image is designed by combining spatial context information, and the energy curve function is defined as follows:

wherein the content of the first and second substances,

is a constant term, ensuring an energy value E_lGreater than 0. As can be seen from equation (1), the gray scale value for a pixel in image I is [ t ]₁,t₂]When l is equal to t₁Here, a two-dimensional binary matrix B is generated_lMatrix B_lThe pixel of the corresponding element in (1) is 1, and when increasing the value of/, only a small number of elements in the matrix change from 1 to-1, so that the energy function will increase, and when increasing to a certain value, the neighbor elements in the matrix change from 1 to-1.

Step two: delineation of pixel second order (N) at spatial locations²) The neighboring pixel (i, j), as shown in fig. 1.

The probability of a certain pixel being noise is calculated based on the energy curve as follows:

wherein, the lambda is an adjusting parameter,

the objective function of the ENFCM may be defined as follows:

wherein:

the function of space restriction is realized,

(d_jkrepresents a sample point x_kAnd x_jEuclidean distance) is a penalty factor representing the sample point x_kSurrounding sample points and x_kSpatial correlation of the classification results, ω_ikThe larger the value of (d), the stronger the spatial correlation.

Step three: according toMethod of FCM minimization, order

The following can be obtained:

from the above formula, it can be seen that in the introduction of the energy curve E_lAnd ω_ikThereafter, the cluster centers are readjusted, the sample membership is indirectly changed by the change of the cluster centers in the iterative process, and E_lAnd ω_ikAll taking into account the sample distribution around the sample point, E_l> 0, and ω_ikOmega when the sample is evenly distributed_ikIs a constant value.

As can be seen from the equations (5) and (6), after the energy curve and the parameter term are directly introduced into the objective function, only the cluster center is constrained. In order to improve the spatial correlation of the membership degree, the membership degree shown in formula (5) is redefined as follows by a weighting method:

wherein

Represents the mean of the degrees of membership.

Drawings

FIG. 1 is a diagram depicting a neighborhood window at a spatial location provided by the present invention.

FIG. 2 is a graph showing the SA variation of the segmentation accuracy of each algorithm in the case of gradual noise enhancement according to the present invention

FIG. 3 is a flow chart of a specific implementation of the ENFCM algorithm provided by the present invention.

Fig. 4 is an abstract drawing provided by the invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and embodiments.

In order to test the anti-noise stability of the ENFCM algorithm, noises (salt and pepper noises and Gaussian noises) with different intensities are added into a flame image of the rotary kiln respectively. In view of the effectiveness of the segmentation accuracy SA in measuring the anti-noise aspect of the algorithm, the anti-noise stability of each algorithm is analyzed by adopting the SA. As can be seen from fig. 2: with the gradual increase of the noise intensity, the segmentation accuracy SA of each algorithm is gradually reduced; since spatial correlation between pixels is not considered and noise is sensitive, SA of the conventional FCM rapidly decreases; FGFCM introduces a histogram method to enable gray levels to replace pixel points, and a correction sample method is adopted to accelerate the operation speed of the algorithm, but the algorithm precision is poor; the SCFMFCM algorithm improves the influence of an image on noise by combining MRF while compressing spatial information by using a histogram, and improves the accuracy of the algorithm; the SCMSFCM and the NDFCM combine pixel space information on the basis of a traditional FCM algorithm, and the segmentation accuracy SA is relatively slow to decline; the ENFCM introduces a space energy curve on the basis of the NDFCM, so that the space correlation of the algorithm is increased, and the segmentation quality of the algorithm is further improved. Therefore, although the noise intensity increase has a certain influence on the algorithm of the invention, the influence is relatively small, which shows that the algorithm has good anti-noise stability. Therefore, the method has a good segmentation effect on the rotary kiln image with high noise degree.

The ENFCM is realized by the following steps:

step 1: initializing parameters, including: the cluster number c (c is more than or equal to 2 and less than or equal to n), an update stop threshold epsilon, a fuzzy weighting index m and the like;

step 2: calculating an energy curve function E according to equation (1)_l；

And step 3: solving a clustering center and a membership degree according to the traditional FCM as the basis of a new algorithm so as to reduce the updating period of the next operation and shorten the operation time;

and 4, step 4: solving for sample x according to traditional FCM algorithm_jAnd cluster center x_kEuclidean function between, calculating ω_ik；

step 5, calculating alpha by using the formula (3)_i ^*Obtaining a sample x_j；

Step 6: calculating u according to equation (7)_ik ^*；

And 7: calculating the clustering center v of the sample class according to equation (6)_ik；

And 8: updating the membership and the cluster center when

The operation is stopped; otherwise, t is t +1, and the process jumps to Step 6.

The specific implementation flow chart of the ENFCM algorithm is shown in fig. 3.

Claims (1)

1. The invention provides an FCM algorithm based on energy noise detection on the basis of NDFCM and is used for realizing initial segmentation of a rotary kiln flame image. The NDFCM algorithm obtains a good anti-noise effect by combining with a spatial neighborhood constraint condition, but the method needs to consider and select a proper window to prevent the loss of image detail information when setting parameters. In order to overcome errors caused by a window in the image segmentation process, the FCM algorithm based on energy noise detection introduces data information of an energy curve compressed image by combining spatial information so as to realize accurate and efficient segmentation of a rotary kiln flame image.

The optimization algorithm comprises the following steps:

the method comprises the following steps: let I ═ l_ijWhere 1. ltoreq. i.ltoreq.m, 1. ltoreq. j.ltoreq.n is an image of size mxn, where l_ijIs the gray value of the image I at pixel location (I, j). L is the maximum gray value of the image I. Defining the d-order neighborhood system of (i, j) as

A second-order neighborhood system is utilized, i.e., (u, v) ∈ { (+ -1, 0), (0, + -1), (1, + -1), (-1, + -1) }. Two-dimensional matrix B generated from I image_l＝{b_ij,1≤i≤m,1≤j≤n}，C＝{c_ijI is more than or equal to 1 and less than or equal to m, j is more than or equal to 1 and less than or equal to n, where c_ijWhen generating the energy curve, different positions are adopted according to a d-order neighborhood system, and the space is combinedThe context information designs an energy curve function associated with each gray value in the image, defined as follows:

wherein the content of the first and second substances,

is a constant term, ensuring an energy value E_lGreater than 0. General formula (1)

It can be seen that for image I, the gray scale value for the pixel is at [ t ]₁,t₂]When l is equal to t₁Here, a two-dimensional binary matrix B is generated_lMatrix B_lThe pixel of the corresponding element in (1) is 1, and when increasing the value of/, only a small number of elements in the matrix change from 1 to-1, so that the energy function will increase, and when increasing to a certain value, the neighbor elements in the matrix change from 1 to-1.

Step two: delineation of pixel second order (N) at spatial locations²) The neighboring pixel (i, j), as shown in fig. 1.

The probability of a certain pixel being noise is calculated based on the energy curve as follows:

wherein, the lambda is an adjusting parameter,

the objective function of the ENFCM may be defined as follows:

wherein:

the function of space restriction is realized,

(d_jkrepresents a sample point x_kAnd x_jEuclidean distance) is a penalty factor representing the sample point x_kSurrounding sample points and x_kSpatial correlation of the classification results, ω_ikThe larger the value of (d), the stronger the spatial correlation.

Step three: a method of minimizing according to FCM, order

The following can be obtained:

from the above formula, it can be seen that in the introduction of the energy curve E_lAnd ω_ikThereafter, the cluster centers are readjusted, the sample membership is indirectly changed by the change of the cluster centers in the iterative process, and E_lAnd ω_ikAll taking into account the sample distribution around the sample point, E_l> 0, and ω_ikOmega when the sample is evenly distributed_ikIs a constant value.

As can be seen from the equations (5) and (6), after the energy curve and the parameter term are directly introduced into the objective function, only the cluster center is constrained. In order to improve the spatial correlation of the membership degree, the membership degree shown in formula (5) is redefined as follows by a weighting method:

wherein

Represents the mean of the degrees of membership.

Images