KR101084719B1 - Intelligent smoke detection system using image processing and computational intelligence - Google Patents

Abstract

PURPOSE: An intelligent smoke detecting system which uses an image processing process and intelligence calculation is provided to detect fire in indoor and outdoor space by simply installing a camera. CONSTITUTION: An intelligent smoke detecting system includes an image input part(10), a data storage part(20), a parameter extraction part(30), a learning part(40), and a determination part(50). The image input part receives image data. The data storage part stores the inputted image data. A parameter extraction part receives the image data from the image input part or data storage part and extracts one or more parameters presenting smoke properties. The learning part receives the parameter from the parameter extraction part and executes learning processes. A learning model is generated by the execution of learning processes. The determination part determines smoke detection by receiving the parameter from the parameter extraction part using the learning model.

Description

Intelligent Smoke Detection System using Image Processing and Computational Intelligence

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a smoke detection system, and more particularly, to an intelligent smoke detection system using image processing and intelligent arithmetic that can detect smoke not only indoors but also outdoors based on a video image.

Conventionally, for fire monitoring, a fire sensor such as a flame sensor, a smoke sensor, or a temperature sensor installed in a ceiling of a structure is used.

If such a fire detector is used, there is a high possibility of malfunction due to a rise in temperature or smoke caused by a cause other than a fire. In addition, even when receiving a signal from the central control station to determine the fire situation it was difficult to determine the exact fire situation only by the signal from the fire detector.

In addition, the fire detector is easy to install in an indoor structure, but it is not suitable for installation in the event of a fire in the outdoors, and even if it is installed, it covers a large outdoor area due to the limitation of the performance of the fire detector. There is a problem that is difficult to do.

In addition, although many studies on fire monitoring apparatus based on image processing using motion vectors have been conducted, it is still difficult to determine the exact fire.

The present invention has an object to solve the above technical problem, to provide an intelligent smoke detection system that can monitor the fire by video processing and intelligent calculation using video image to determine the exact fire situation There is that purpose.

Another object of the present invention is to provide an intelligent smoke detection system capable of monitoring a fire not only indoors but also outdoors by installing a simple camera.

An intelligent smoke detection system using image processing and intelligent operation according to an embodiment of the present invention includes an image input unit for receiving image data; A data storage unit for storing the received image data; A parameter extraction unit configured to receive image data from the data storage unit or the image input unit and extract one or more parameters representing the characteristics of the acting; Learning unit for generating a; And a determination unit that receives a parameter from the parameter extraction unit and determines whether smoke is detected using the learning model.

Specifically, the at least one parameter extracted by the parameter extractor may include a disorder measuring the degree to which the appearance of the subject changes randomly; An increase that means the rate of change of the target size based on the diffusion characteristics of the smoke; A movement frequency for calculating how much movement has occurred in the image of the image data; A local wavelet energy gradient of smoke calculated from values of each pixel of the image data; And a magnetic similarity in which the shape of the entire region of the image data is repeatedly displayed in the portion of the region.

In addition, the disorder diagram detects a moving pixel of the image data and converts it into a binarized image, obtains a cluster of pixels moved by the binarized image, and divides the pixel estimated to be smoke into the cluster, and thus the ratio of the estimated smoke pixels. It is characterized by calculating by calculating.

Specifically, the moving pixel detection method uses a background removal algorithm, and the background removal algorithm selects and removes pixels corresponding to a background when there is no change between frames in the input image or when the change is common. It is characterized by.

In addition, the increase is characterized in that calculated by the following equation.

(Where t is time or frame, R is target area of target frame, and Size is area, respectively.)

In detail, the motion frequency is obtained by extracting only a moving region from the image data, converting the image into a gray image, converting the image into a wavelet format, and passing a high pass filter and a low pass filter, respectively, to obtain a high frequency image and a low frequency image, respectively. It is preferable to calculate the sum of each of the and low frequency images by obtaining a predetermined frame to obtain a high frequency accumulated image and a low frequency accumulated image, and to calculate the high frequency accumulated image by dividing the high frequency accumulated image by the low frequency accumulated image.

In more detail, after obtaining the high frequency image and the low frequency image, the high frequency image and the low frequency image are downsampled, respectively, in order to increase the sharpness of each image.

The local wavelet energy gradient may include (a-1) converting the image data and the background data into a gray image; (a-2) converting the gray images into a wavelet format and acquiring three bands of images by two-step filtering; (a-3) generating a new composite image of the image data and the background data by squaring and combining respective pixel values of the three band images; (a-4) dividing the newly synthesized images into n blocks each; (a-5) obtaining n energy values en for divided n blocks of the composite image of the image data; (a-6) obtaining n energy values ebn for the n divided blocks of the composite image of the background data; (a-7) setting a postponing region where ebn is greater than a predetermined predetermined value than en; And (a-8) obtaining a difference between the energy of the background data and the energy of the image data for each of the n blocks with respect to the suspected smoke area, and dividing the energy of the background data by the energy of the background data to calculate an energy gradient. It is preferable to calculate by step.

Specifically, in the step (a-2), the three band images pass the high pass filter and then pass the high pass filter again, the high pass filter passes the low pass filter, and the low pass image. After passing the filter is characterized in that the image passed the high pass filter.

In addition, the magnetic similarity is calculated by selecting an average value of the selected values after selecting the series of the image data so as not to overlap with a predetermined predetermined size.

Specifically, when the auto-similarity uses a time-related series, it is calculated by using a Hurst parameter representing a decay rate of the autocorrelation function of the series using an autocorrelation model.

The neural network for generating the learning model uses a fuzzy neural network, the modeling of the fuzzy system uses an IF-THEN rule, and the structure of the fuzzy neural network is preferably determined using a clustering technique.

Specifically, the structure of the fuzzy neural network is determined by the central value v _i , the number of clusters u _ik and the weight w _i .

In addition, the structure of the fuzzy neural network is determined by initializing each variable including the weight value m of the exponent (b-1), the maximum number of clusters C, the error tolerance ε, the maximum number of repetitions T, and the initial validity value Z (c). step; (b-2) initializing the number of clusters u _ik (t) to 0 or 1; (b-3) calculating a new center value v _i (t) and a weight w _i (t); (b-4) updating the number u _ik (t) of the clusters using the calculated new center value v _i (t) and the weight w _i (t); (b-5) calculating a maximum value Delta of the absolute value of the difference between u _ik (t) and u _ik (t-1); (b-6) if the maximum value Delta is greater than the error tolerance ε, increase t by one, and repeating the steps after step (b-3); (b-7) calculating a new validity value Z (c) if the maximum value Delta is less than or equal to the error tolerance ε; (b-8) if Z (c) is greater than Z (c-1), ending; And (b-9) if Z (c) is less than or equal to Z (c-1), increasing c and repeating the steps after step (b-2); As determined by c, the number means the number of clusters, and t means the number of repetitions.

In addition, the learning model generation according to an embodiment of the present invention is characterized in that asymmetric Gaussian membership function tuning using a genetic algorithm.

The method of tuning comprises the steps of: (c-1) calculating initial values for weights and variances; (c-2) copying the initial weights and variances to the last genotype and generating the initial gene population by generating random values for the remaining genotypes; (c-3) calculating the goodness of fit for the initial gene population using fuzzy inference and an error function; (c-4) generating a next generation from genetic algorithm operator and selected individuals; (c-5) calculating the goodness of fit for the newly created gene population using fuzzy inference and an error function; And (c-6) if the optimal gene is found and terminates, or repeatedly performing the step after the step (c-4).

Specifically, the step (c-4) is a step (c-4-1) performing a selection operator to place the genotype having the best fit to the last genotype by comparing the goodness of each genotype with the goodness of the last genotype; (c-4-2) performing a crossover operator to cross over genotypes whose survival is determined by selection; And (c-4-3) optionally performing a mutation operator for selecting and mutating genotypes.

According to the present invention, it is possible to provide an intelligent smoke detection system capable of monitoring a fire by image processing and intelligent calculation using a video image so as to grasp an accurate fire situation.

In addition, according to the present invention it is possible to provide an intelligent smoke detection system capable of monitoring the fire not only indoors, but also outdoors by installing a simple camera.

1 illustrates an intelligent smoke detection system using image processing and intelligent calculation according to an exemplary embodiment of the present invention.
Figure 2 shows the result of comparing the disorder for the walker (Walker) and smoke (Smoke).
3 illustrates an increase in the walking and acting of the ordinary person in the video, respectively.
4 illustrates a process of obtaining a high frequency image (H) and a low frequency image (L) by passing a high pass filter (HPF) and a low pass filter (LPF), and performing downsampling.
5 shows the frequency of movement.
6 is an example of knowing the local wavelet energy gradient.
7 shows a general fuzzy neural network system structure for smoke detection.
8 is a flowchart for determining a system structure and a number of rules according to an embodiment of the present invention.
9 shows a flowchart for tuning of an asymmetric Gaussian membership function.
10 shows a fuzzy neural network system structure for intelligent smoke detection to which smoke features are applied.

Hereinafter, an intelligent smoke detection system according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The following examples of the present invention are intended to embody the present invention, but not to limit or limit the scope of the present invention. It will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

First, FIG. 1 shows an intelligent smoke detection system using image processing and intelligent calculation according to an embodiment of the present invention.

As can be seen from Figure 1 intelligent smoke detection system according to an embodiment of the present invention comprises an image input unit 10; A data storage unit 20; Parameter extraction unit 30; Learning unit 40; And a determination unit 50;

Hereinafter, the role of each component will be described in detail.

The image input unit 10 receives an image obtained by capturing a moving image of a target area requiring smoke detection for monitoring a fire by using a device capable of capturing a moving image such as a CCD camera.

The data storage unit 20 stores the received image data.

The parameter extraction unit 30 receives image data from the data storage unit 20 or the image input unit 10 directly, and extracts one or more parameters representing the characteristics of smoke using the input image data. do.

The learner 40 receives parameters from the parameter extractor 30 to perform learning, and generates a learning model by performing learning, and the determiner 50 uses the learning model. By receiving the parameter from the parameter extraction unit 30 serves to determine whether the smoke is detected.

The parameter extraction unit 30 will be described in more detail.

The parameter extracted by the parameter extraction unit 30 is a feature for detecting smoke and is composed of those which can show a large difference in values between the smoke and other objects. Five features are presented in the present invention.

That is, the parameter extractor 30 may measure a degree of disorder in the appearance of the object in order, Disorder, an increase in meaning of a change rate of the size of the object based on the diffusion of smoke, and the image. Frequent Flicker, which calculates how much movement has occurred in the image of the data, Local Wavelet Energy of Smoke calculated from the values of each pixel of the image data, and the image data. One or more parameters are selected and extracted from a set of parameters representing the characteristics of the smoke including the self-similarity that is repeated in the region of the region.

The characteristics of each parameter will be described in more detail.

Disorder

Disorder is a measure of the degree of disorder in the appearance of an object, and acting does not have a regularized form, compared to other objects, and changes in disorder.

In order to measure the disorder, first, the input signal is processed through a gray and median smooth filter (S11). Only moving pixels are detected using the background elimination algorithm (S12). The image is converted into a binarized image (S13), and a cluster of moving pixels is obtained by removing noise through a morphology filter (S14). Based on the clustering of these pixels, pixels similar to smoke are determined from the original image. Here, the size of the cluster is used as the disorder detection data. The ratio of the estimated smoke pixels is obtained by dividing the pixel estimated as smoke by the group of pixels obtained from the binarized image (S15).

The background elimination algorithm used in the step S12 is characterized in that if there is no change between frames in the input image or if the change is ordinary, it is determined as a pixel corresponding to the background, and then selected and removed.

Figure 2 shows the result of comparing the disorder for the walker (Walker) and smoke (Smoke).

Growth

The increase is a measure of how the size of the object has changed. Smoke has a characteristic of being generally diffused compared to other objects. Based on these characteristics, the rate of change of the size of the object is checked. Equation 1 is used for this.

Here, t represents time or a frame, and R represents a target frame target region. In the measurement method, the input signal is first processed by gray and median smooth filters. The background removal algorithm is used to detect only moving pixels. The image is transformed into a binarized image and the morphology filter removes the noise to obtain a cluster of moving pixels. The size of this cluster is used as the increase detection data. The increase is measured by comparing the size values obtained at regular intervals.

3 shows the increase in the walking and acting of the ordinary person from the video. It can be seen that the value for acting is greater than the value of the person walking.

Frequent Flicker

The frequency of motion is how many motions occur in the image. The characteristic of the smoke is that it is affected by the wind, which causes more movement than other objects. The method of calculating the motion frequency first converts the image from which only the moving region is extracted into a gray image (S21), and then converts the image into a wavelet format (S22). As can be seen from FIG. 4, the high pass filter (HPF) and the low pass filter (LPF) are passed through only one stage to obtain a high frequency image (H) and a low frequency image (L). The high frequency image and the low frequency image can be down sampled respectively (S24), but the reason for down sampling is that each low frequency or high frequency component appears more clearly as down sampling occurs.

As shown in Equation 2, a high frequency cumulative image and a low frequency cumulative image of accumulating a predetermined frame are generated (S25), and a value obtained by dividing the high frequency cumulative image by the low frequency cumulative image is adopted as a motion frequency characteristic value in the present invention.

When the graph is drawn by the adopted motion frequency characteristic value, it is shown in FIG.

As can be seen from FIG. 5, it can be seen that the moving person's moving picture is very severe and the movement frequency characteristic value is also maintained higher than the moving picture. The lower blue and red portions of FIG. 5 represent values of the smoke video, and the purple and green colors above are values of the general video. When comparing the value of the acting video and the value of the general video through Figure 5 it can be seen that there are a lot of significant differences.

The principle of the above-mentioned process of calculating the frequency of motion will be described. The image passing through the high pass filter is a high frequency component of the image, that is, an image having only an object or a boundary part of the background. On the other hand, the image passing through the low pass filter is an image having a portion other than the boundary portion. By accumulating these two images for a certain frame and then finding the rain, we can figure out how often the boundary has moved.

Local Wavelet Energy

When smoke covers an image, it appears to be blurry as seen through thin paper. The value of each pixel using this feature is determined as energy, and the change is observed to obtain a value. First, in order to obtain energy, the original image data and the background data are converted into gray images, respectively (S31), and then, through the two-stage filter using a wavelet, images of three bands are obtained (S32).

The background data of step S31 is obtained by a background removal algorithm.

Looking at the process of acquiring the images of the three bands, after passing the high pass filter and passing the high pass filter again to obtain each of the HH band image (S321), after passing the high pass filter and low pass filter Passing through the step of obtaining the image of the HL band (S322) and passing through the low pass filter, and then passing through the high pass filter to obtain the image of the LH band, respectively (S323).

 By combining squares of the pixel values of the three band images, new composite images of the image data and the background data, that is, Wn and Wbn, are generated (S33).

That is, for example, Wn may be calculated as in Equation 3.

Next, the newly synthesized images are divided into n blocks (S34). N energy values en are calculated for the n divided blocks of the composite image of the image data (S35). Further, n energy values ebn are calculated for the n divided blocks of the composite image of the background data (S36).

For example, when the synthesized image is divided into blocks of 8 * 8 size, equations 4 and 5 may be used to calculate energy values from each block. En is obtained from the video of Wn and Ebn is obtained from the video of Wbn.

An index is assigned to each block for both Wn and Wbn based on the energy values calculated as described above. Satisfying the condition of Equation 6 for each index is indicated in red in FIG. 6.

In FIG. 6, the region marked with red is considered to be a region where energy changes, that is, smoke is suspected. In other words, the suspected area where the ebn is greater than a predetermined predetermined value than the en is set (S37).

The energy change degree of the suspected smoke region may be calculated by Equation 7 (S38).

The range of suspected smoke is quite large, so the rate of change in each suspected area is calculated and averaged to yield meaningful energy characteristic values. Since the change in the energy value of the image where the smoke appeared is very large, it is easy to distinguish the presence or absence of the smoke by comparing the image before the smoke and the image after the smoke.

Self-Similarity

Magnetic similarity refers to a property in which the shape of the whole area is repeated in the area part. The smoke zone has the property of self-similarity within the smoke zone. In other words, each part of the smoke area is spread or reduced in a similar manner to the movement pattern of the entire area. This coefficient of magnetic similarity is detected.

라 정의하자. In other words, the self-similarity refers to the characteristic that the new series has the same autocorrelation function as the original series even when multiple series are multiplexed. Of data

Let's define

는 수학식 8로 표현된다.The magnetic similarity is obtained by selecting values such that the series of original data X does not overlap by m, and then averages the selected values. If X has autosimilarity, then X's autocorrelation function

Is expressed by Equation (8).

는 모든 m에 대한 급수

에 대해서도 같은 자기 상관 함수를 가지며, 이와 같은 특성은 급수들이 분포적인 자기 상관성(distributional self-similarity)을 가진다고 한다.

Watering for all m

We have the same autocorrelation function for, and this property is said that the series have distributed self-similarity.

An autocorrelation model is used to represent time-related series, because the autocorrelation model can be used to express the degree of autocorrelation of a series using a single parameter. This parameter represents the decay rate of the autocorrelation function of the series, which is referred to as the Hurst parameter.

To analyze the self-similarity feature, we can find the Hurst parameter defined above.

System configuration for intelligent smoke detection

In the present invention, the learning model was designed in consideration of the following points in order to maximize the learning and recognition effects.

How much learning do you need?

How can I partition the data into test and validation data?

What are the defuzzy methods, error functions, and evaluation functions that can be used?

How can I select important input variables?

Can the neural network be used in safety critical applications?

The present invention proposes a new learning method that performs membership function tuning to find an optimal solution.

The fuzzy clustering method is used to determine the initial fuzzy membership function, that is, the structure of the system, and to optimize the membership function using the genetic algorithm to determine the optimal initial value. Genetic algorithms can be used to solve adaptive learning and optimization problems by combining neural networks and fuzzy logic in terms of the following characteristics.

-Multiple local solutions exist

-Some degree of regularity

-Problem areas can be expressed to some degree in regularity

Fuzzy Neural Network System Modeling for Intelligent Smoke Detection

For systems with multiple inputs and one output, the dimensions of the input variables and outputs are fixed in advance according to the given system. The model used in the present invention uses a simplified fuzzy model. In computational intelligence, fuzzy neural network structure is basically an inferential system composed of neural network structure. When n- dimensional vectors are applied to the fuzzy inference system, fuzzy rules are created in the learning phase. When the input variable is represented by x = ( x ₁ , x ₂ , ......., x _n ) and the output value is represented by y , the " IF-THEN " rule of the form of Equation 10 is used for fuzzy inference. Use In the " IF-THEN " rule, the front part from the IF to the front part of the THEN is called the front part, and the lower part of the IFN is called the back part.

In the above, the index i represents the order of rules, and A _ij represents the i th rule and the j th language variable. w _i gets the real value corresponding to the weight for calculating the output at the conclusion.

In the present invention, a Gaussian function represented by Equation 11 is used as a membership function for fuzzy neural networks.

Fuzzy inference is required for knowledge base application, and Equation 12 is used to calculate the belonging value of the i th rule for the input vector.

The values of membership and weight for the premise of the rule are still fuzzy values. Therefore, in order to obtain the final result for a given input value, it is defused using Equation 13 and then used as the final value.

To build a better system using the strengths of fuzzy set theory and neural network theory to solve a given problem, we need to learn from learning, optimization, connection structure, human - friendly " IF-THEN " rules, and knowledge from experts. Must be able to obtain Using them, the fuzzy neural network system can be expressed as a fuzzy inference system through the concatenation of the fuzzy " IF-THEN " rule using the predicate part and the result part. This structure allows you to perform learning and control the rules autonomously.

Such a system uses the training data to generate fuzzy rules in the learning stage, thereby producing better results than conventional methods. The basic idea is the clustering method to determine the fuzzy rules. The structure of the proposed system is composed of an input layer, a hidden layer, and an output layer, as shown in FIG.

Structural Decision of Fuzzy Neural Network System for Intelligent Smoke Detection

로 정의했을 때 수학식 15는 목적 함수를 나타낸다.Fuzzy c-Means clustering technique was used to understand the structure of the system. In this case, both weights and desired output information are used as shown in Equation (14). N pairs of data

When defined as 15 represents the objective function.

Where v _i is the vector with the center of the i th rule in the input space and w _i is the output of the i th rule. u _ik has a value between 0 and 1 depending on the degree to which the k th input data x _k belongs to the i th cluster. m is usually used as the weight of the exponent which indicates the effect on the degree of fuzzyness (ambiguity) of the membership function.

Since the center value and the weight are fixed, the membership function can be obtained using Equation 16.

After the membership function is calculated, the center value v _i and the weight w _i can be calculated using Equations 17 and 18.

The algorithm for determining the system structure and the number of rules will be described in more detail below with reference to FIG. 8.

Algorithms for Determining System Structure and Number of Rules

First, each variable including the weight value m of the exponent, the maximum number of clusters C, the error tolerance ε, the maximum number of repetitions T, and the initial validity value Z (c) is initialized (S41). In addition, the number of clusters u _ik (t) is initialized to 0 or 1 (S42).

Next, the new center value v _i (t) and the weight w _i (t) are calculated (S43). The number of clusters u _ik (t) is updated using the calculated new center value v _i (t) and the weight w _i (t) (S44).

Calculate the maximum value Delta of the absolute value of the difference between u _ik (t) and u _ik (t-1) (S45), and if the maximum value Delta is greater than the error tolerance ε, increase t by one, after step S43 Repeat the steps of (S46. S47).

If the maximum value Delta is less than or equal to the error tolerance ε, a new validity value Z (c) is calculated (S48). If Z (c) is larger than Z (c-1), the algorithm ends (S49, S51).

However, if Z (c) is less than or equal to Z (c-1), c is increased (S49, S50), and the steps after the step S42 are repeatedly performed.

Initialization values of step S41 are as follows.

Exponential weighting: m> 0

Error Tolerance (Threshold): ε> 0

Maximum number of clusters: C

Maximum number of repetitions: T

Initial cluster validity value: Z (1)

을 사용하여 산출된다. 즉, d_ik는 각 그룹 내에서의 에러(WGSS-Within Group Sum of Squared Error)를 나타내며, dd_io은 그룹 사이의 에러(BGSS- Between Groups Sum of Squared Error)를 나타내므로, Z(c)의 최소화가 군집 수를 선택하는 기준으로 사용되는 것이다.
However, the validity value in the algorithm for determining the system structure and the number of rules is

Is calculated using That is, d _ik denotes an error within each group (WGSS-Within Group Sum of Squared Error), and dd _io denotes an error between groups (BGSS- Between Groups Sum of Squared Error). Minimization is used as a criterion for selecting the number of clusters.

Parameter Tuning of Fuzzy Membership Function Using Genetic Algorithm

A method for optimizing the parameters of the Gaussian membership function used in Equation 11 using a genetic algorithm will be described. In the present invention, since the asymmetric membership function is proposed, the variance is divided into left and right sides, and the genetic algorithm is used to optimize these parameters.

The defuzzy method for the output of the fuzzy system uses Equation 13 and defines an error function as Equation 19 to evaluate the goodness of fit.

로 표현할 수 있고, 입력과 그에 대한 목표 값의 쌍은

로 표현한다. 이를 바탕으로 각 규칙에 대한 초기 중심 값과 분산 값을 수학식 20과 수학식 21을 이용한다.In the previous stage of identifying system structure, the number of optimized fuzzy rules was obtained. kth input data

Where the input and the target value for it are

Expressed as Based on this, equations (20) and (21) are used as initial center values and variance values for each rule.

In the present invention, when the parameter values are tuned using a genetic algorithm, the parameter values are composed of values different from the initial values. In general genetic algorithm, the initial value is assigned an arbitrary value, and due to the characteristics of the system, the real gene is used within a certain range. The central values for each rule are obtained at the system architecture level. These values define that they do not need to be changed because they are representative of each cluster during the clustering stage.

Also, the parameters related to variance are tuned within a certain range so as not to deviate from the range of associated central values. However, the range of parameter values related to the weight value is not particularly limited. The equation for the membership function in the fuzzy rules changes Equation 11 to Equation 22 for the asymmetric membership function. Therefore, using this equation, it can be seen that it becomes an asymmetric membership function using left and right variance values.

The number of chromosomes to be tuned in each cluster depends on the number of elements in the input data. For example, in the system of the present invention, when the input is configured in five dimensions for each rule, the system includes 11 parameters including one parameter related to 10 left and right variances and one weight parameter.

Once the system structure has been determined, the tuning of the asymmetric Gaussian membership function will next be described according to the flow chart of FIG.

Tuning Asymmetric Gaussian Membership Function Using Genetic Algorithm

First, an initial weight value and an initial value for variance are calculated using Equations 20 and 21 (S61).

Next, the initial weight value and the variance value are copied to the last genotype, and the remaining genotypes generate random values to generate an initial gene population (S62).

The fitness for the initial gene population is also calculated using fuzzy inference and an error function (Equation 19) (S63). Next, reproduction, that is, the next generation is generated from the genetic algorithm operator and the selected individuals (S64).

A fitness for the newly generated gene population is also calculated using fuzzy inference and an error function (Equation 19) (S65). If the optimal gene is found, it ends. Otherwise, go to step S64 and repeat.

The step S64 may include performing a selection operator (S641) comparing the goodness of fit of each genotype to the goodness of the last genotype and placing the genotype having the best fit in the last genotype (S641); Performing a crossover operator crossing the genotype whose survival is determined by selection (S642); And optionally performing a mutation operator (S643) for selecting and mutating genotypes.

Figure 10 shows a structure that can be applied to the fuzzy neural network using the characteristics of the smoke proposed in the present invention. This structure shows a structure similar to a multilayer omnidirectional form in neural networks. The input layer uses the detected smoke characteristics as input elements. The hidden layer has a value between 0 and 1 depending on the extent of each cluster according to the value of the input data. w represents a weight and is determined in the learning process. y means the result value.

Using the learning model having the structure as shown in FIG. 10 generated by the above-described process, the determination unit 50 determines whether the smoke to be detected in the monitoring target area that needs to detect the actual smoke.

10: video input unit 20: data storage unit
30: parameter extraction unit 40: learning unit
50: judgment unit

Claims (21)

In the intelligent smoke detection system,
An image input unit configured to receive image data;
A data storage unit for storing the received image data;
A parameter extracting unit which receives image data from the data storage unit or the image input unit and extracts at least one parameter representing a feature of smoke;
A learning unit which receives a parameter from the parameter extracting unit to perform learning and generates a learning model by performing learning; And
And a determination unit that receives a parameter from the parameter extraction unit to determine whether to detect smoke using the learning model. The intelligent smoke detection system using image processing and intelligent operation.
The method of claim 1, wherein the one or more parameters extracted by the parameter extractor are:
Disorder to measure the degree to which the appearance of the subject changes disorderly;
An increase that means the rate of change of the target size based on the diffusion characteristics of the smoke;
A movement frequency for calculating how much movement has occurred in the image of the image data;
A local wavelet energy gradient of smoke calculated from values of each pixel of the image data; And
Intelligent similarity detection system using image processing and intelligent operation, characterized in that the shape of the entire area of the image data is selected from a set of parameters including;
The method of claim 2, wherein the disorder is
Is calculated by detecting moving pixels of the image data and converting them into a binarized image, obtaining a cluster of pixels moved by the binarized image, and dividing the pixels estimated to be smoke by the clusters to obtain a ratio of the estimated smoke pixels. Intelligent smoke detection system using image processing and intelligent operation.
The method of claim 3,
The moving pixel detection method uses a background removal algorithm, wherein the background removal algorithm selects and removes pixels corresponding to a background when there is no change between frames in the input image or when the change is common. Intelligent smoke detection system using image processing and intelligent computation.
The intelligent smoke detection system according to claim 2, wherein the increase is calculated by the following equation.

(Where t is time or frame, R is target area of target frame, and Size is area, respectively.)
The method of claim 2, wherein the frequency of movement,
After extracting only the moving region from the image data, converting it into a gray image, converting it into a wavelet format, passing a high pass filter and a low pass filter, respectively, to obtain a high frequency image and a low frequency image, respectively, and the high frequency image and the low frequency image are respectively fixed. Accumulating frames to obtain a sum thereof to obtain a high frequency accumulated image and a low frequency accumulated image, and calculating the high frequency accumulated image by dividing the high frequency accumulated image by the low frequency accumulated image.
The method of claim 6, wherein after obtaining the high frequency image and the low frequency image, down-sampling the high frequency image and the low frequency image to increase the sharpness of each image is performed. Detection system.
The method of claim 2, wherein the local wavelet energy gradient is,
(a-1) converting the image data and the background data into a gray image;
(a-2) converting the gray images into a wavelet format and acquiring three bands of images by two-step filtering;
(a-3) generating a new composite image of the image data and the background data by squaring and combining respective pixel values of the three band images;
(a-4) dividing the newly synthesized images into n blocks each;
(a-5) obtaining n energy values en for divided n blocks of the composite image of the image data;
(a-6) obtaining n energy values ebn for the n divided blocks of the composite image of the background data;
(a-7) setting a postponing region where ebn is greater than a predetermined predetermined value than en; And
(a-8) obtaining a difference between the energy of the background data and the energy of the image data for each of the n blocks with respect to the suspected smoke area, and dividing the energy by the energy of the background data to calculate an energy change degree; Intelligent smoke detection system using image processing and intelligent operation, characterized in that calculated by.
The method of claim 8, wherein the three bands of the step (a-2),
After passing through the high pass filter, the image passed the high pass filter again, passed through the high pass filter, passed the low pass filter image, and passed through the low pass filter, characterized in that the image passed through the high pass filter Intelligent smoke detection system using image processing and intelligent operation.
The method of claim 2, wherein the magnetic similarity is
Intelligent smoke detection system using image processing and intelligent operation, characterized in that calculated by the average value of the selected value after selecting the water supply of the image data is not overlapped to a predetermined predetermined size.
The method of claim 10,
When the auto-similarity uses a time-related series, the intelligent smoke detection system using the image processing and intelligent calculation, characterized in that it is calculated by the Hurst parameter representing the decay rate of the auto-correlation function of the series.
The method of claim 1,
The neural network for generating the learning model uses a fuzzy neural network, the modeling of the fuzzy system uses IF-THEN rules, and the structure of the fuzzy neural network is determined using a clustering technique. Intelligent smoke detection system using computation.
The method of claim 12, wherein the determination of the structure of the fuzzy neural network,
Intelligent smoke detection system using image processing and intelligent arithmetic, characterized in that determined by the center value v _i , the number of clusters u _ik and the weight w _i .
The method of claim 13, wherein the determination of the structure of the fuzzy neural network,
(b-1) initializing each variable comprising a weight m of the exponent, the maximum number of clusters C, the error tolerance ε, the maximum number of iterations T, and the initial validity value Z (c);
(b-2) initializing the number of clusters u _ik (t) to 0 or 1;
(b-3) calculating a new center value v _i (t) and a weight w _i (t);
(b-4) updating the number u _ik (t) of the clusters using the calculated new center value v _i (t) and the weight w _i (t);
(b-5) calculating a maximum value Delta of the absolute value of the difference between u _ik (t) and u _ik (t-1);
(b-6) if the maximum value Delta is greater than the error tolerance ε, increase t by one, and repeating the steps after step (b-3);
(b-7) calculating a new validity value Z (c) if the maximum value Delta is less than or equal to the error tolerance ε;
(b-8) if Z (c) is greater than Z (c-1), ending; And
(b-9) if Z (c) is less than or equal to Z (c-1), increasing c and repeating the steps after step (b-2); Decided,
C denotes the number of clusters, and t denotes the number of repetitions; and intelligent smoke detection system using image processing and intelligent operation.
delete delete 15. The intelligent smoke detection system according to claim 13 or 14, wherein the asymmetric Gaussian membership function is tuned using a genetic algorithm.
The method of claim 17, wherein the tuning method,
(c-1) calculating initial values for weights and variances;
(c-2) copying the initial weights and variances to the last genotype and generating the initial gene population by generating random values for the remaining genotypes;
(c-3) calculating the goodness of fit for the initial gene population using fuzzy inference and an error function;
(c-4) generating a next generation from genetic algorithm operator and selected individuals;
(c-5) calculating the goodness of fit for the newly created gene population using fuzzy inference and an error function; And
(c-6) if the optimal gene is found, terminate the step or repeat the step (c-4) and then repeat the step; intelligent smoke detection system using image processing and intelligent operation, characterized in that it comprises a. .
The method of claim 18, wherein step (c-4),
(c-4-1) performing a selection operator to compare the goodness of fit of each genotype with the goodness of fit of the last genotype and to place the genotype having the best fit into the last genotype;
(c-4-2) performing a crossover operator to cross over genotypes whose survival is determined by selection; And
(c-4-3) performing a mutation operator for randomly selecting and mutating genotypes; and intelligent smoke detection system using image processing and intelligent operation.
20. The intelligent smoke detection system according to claim 19, wherein the initial weights and variances are respectively calculated by the following equations.

(In the above, the k th input data is

Where the input and the target value for it are

Expressed as
20. The method of claim 19,
The error function is obtained by the following equation, Intelligent smoke detection system using image processing and intelligent operation.

(In the above, the k th input data is

Where the input and the target value for it are

Expressed as

Images