KR102231160B1 - Intelligent fire identification apparatus, intelligent fire identification method and recording medium - Google Patents

Abstract

Disclosed are an intelligent fire detection device, an intelligent fire detection method, and a recording medium that extract statistical texture features from infrared thermal image information and analyze them through a deep learning model to calculate probability values such as smoke, fire, and heat reflection. The fire detection apparatus according to an embodiment of the present invention statistically analyzes pixel values from infrared thermal image information to extract primary statistical texture features, and quantitatively analyzes a correlation between pixels in the infrared thermal image information. A texture feature extraction unit for extracting secondary statistical texture features; And a smoke class, a fire class, a smoke heat reflection class, a fire heat reflection class, and a simple high temperature object class by analyzing the first statistical texture features and the second statistical texture features through a deep learning model. It includes a deep learning classifier that calculates probability values.

Description

Intelligent fire identification apparatus, intelligent fire identification method and recording medium}

The present invention relates to an intelligent fire detection device, an intelligent fire detection method, and a recording medium, and more particularly, extracts statistical texture features from infrared thermal image information and analyzes it through a deep learning model, such as smoke, fire, and heat reflection. It relates to an intelligent fire detection device for calculating probability values, an intelligent fire detection method, and a recording medium.

Smart buildings with improved fire monitoring functions are expanding by applying the 4th industrial revolution technology. Existing fire monitoring systems largely include a method based on sensor signals such as ultraviolet rays, a method based on RGB image information such as CCTV, and a method based on infrared thermal images. In the case of a fire monitoring system based on a sensor signal, when the measured signal value exceeds the allowable value (threshold value), it is recognized as a fire. While it is simple and inexpensive, the area that can be monitored is narrow and there is a possibility of a false alarm. There is a high problem with this. In the case of a fire monitoring system based on RGB image information, it is suitable for the purpose of detecting forest fires due to its wide area that can be monitored, but there is a problem that the visibility is limited due to smoke when an indoor fire occurs, so in situations full of smoke appearing after the initial stage of fire occurrence, There is a limit that prevents it from performing the normal monitoring role. In the case of infrared thermal imaging system, due to the advantage of the RGB image information system with a wide monitorable area and the characteristic of transmitting smoke, it can perform a monitoring role in the entire process from the beginning to the end of the fire in a smoke-filled indoor fire. However, there is still a problem in that fire is recognized based on pixel values, and fire, smoke, heat reflection of fire and smoke, and simple high-temperature objects are not accurately recognized.

The present invention extracts statistical texture features from infrared thermal image information, applies a deep learning model, calculates probability values such as smoke, fire, and heat reflection, and detects a fire in a smoke-filled indoor fire situation and estimates the fire source. It is to provide an intelligent fire detection device, an intelligent fire detection method, and a recording medium that enable firefighters to accurately induce fire sources and quickly extinguish fires.

The problem to be solved by the present invention is not limited to the problems mentioned above. Other technical problems not mentioned will be clearly understood by those of ordinary skill in the art from the following description.

The intelligent fire detection apparatus according to an aspect of the present invention statistically analyzes pixel values from infrared thermal image information to extract primary statistical texture features, and quantitatively calculates correlation between pixels from the infrared thermal image information. A texture feature extraction unit that analyzes and extracts secondary statistical texture features; And a smoke class, a fire class, a smoke heat reflection class, a fire heat reflection class, and a simple high temperature object class by analyzing the first statistical texture features and the second statistical texture features through a deep learning model. It may include a deep learning classifier that calculates probability values.

The primary statistical texture features may include at least two or more of an average value, a variance value, a deviation value, asymmetry, and kurtosis of the pixel values. The second-order statistical texture features may include at least two or more of dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and inverse difference between the pixels.

The input layer of the deep learning model may include 12 input nodes consisting of the first statistical texture features and the second statistical texture features. The output layer of the deep learning model may include five output nodes composed of the classes. The deep learning model may have four hidden layers between the input layer and the output layer.

The texture feature extraction unit 1 for selecting the primary statistical texture features from the average value, variance value, deviation value, asymmetry, and kurtosis of the pixel values based on the material and arrangement structure of the door, wall, and window of the fire monitoring area. Tea texture feature selector; And based on environmental information including temperature and humidity of the fire monitoring area, the secondary statistical texture features are selected from among the dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and inverse differential inertia between the pixels. It may include a second texture feature selection unit to select.

The texture feature extractor may include a first texture feature extractor that extracts only a first statistical texture feature selected by the first texture feature selector among the first statistical texture features from the infrared thermal image information; And a second texture feature extracting unit that extracts only a second statistical texture feature selected by the second texture feature selector among the second statistical texture features from the infrared thermal image information.

The deep learning model may include a first step deep learning model and a second step deep learning model. The first input layer of the first-stage deep learning model consists of five input values including the first statistical texture features, and the first output layer of the first-stage deep learning model is a non-critical class corresponding to a simple high-temperature object and Consisting of two output values of an important class excluding the non-critical class, the first-stage deep learning model may have at least three hidden layers between the first input layer and the second output layer. The second input layer of the second-stage deep learning model is composed of eight input values including the important class and seven second-order statistical texture features, and the second output layer of the second-stage deep learning model is fire, smoke, and It consists of four output values corresponding to the non-heat reflection and smoke heat reflection classes, and the second-stage deep learning model may have at least four hidden layers between the second input layer and the second output layer.

According to another aspect of the present invention, there is provided a method comprising: statistically analyzing pixel values from infrared thermal image information to extract primary statistical texture features; Extracting secondary statistical texture features by quantitatively analyzing a correlation between pixels from the infrared thermal image information; And, by analyzing the first statistical texture features and the second statistical texture features through a deep learning model, classes including a smoke class, a fire class, a smoke heat reflection class, a fire heat reflection class, and a simple high-temperature object class. An intelligent fire detection method comprising calculating probability values is provided.

The step of extracting the primary statistical texture features may include a first order including an average value, a variance value, a deviation value, asymmetry and kurtosis of the pixel values, based on the material and arrangement structure of the door, wall, and window of the fire monitoring area. Selecting one or more primary statistical texture features from among the statistical texture features; And, from the infrared thermal image information, it may include extracting only the one or more first-order statistical texture features selected from among the first-order statistical texture features.

The extracting of the secondary statistical texture features may include dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and Selecting one or more second-order statistical texture features from among second-order statistical texture features including an inverse degree of inertia; And, from the infrared thermal image information, it may include extracting only the one or more secondary statistical texture features selected from among the secondary statistical texture features.

According to another aspect of the present invention, a computer-readable recording medium in which a program for executing the intelligent fire detection method is recorded is provided.

According to an embodiment of the present invention, an intelligent fire detection device that extracts statistical texture features from infrared thermal image information and analyzes it through a deep learning model to calculate probability values such as smoke, fire, and heat reflection, and an intelligent fire detection method and recording. Media is provided.

In addition, the present invention identifies smoke, fire, smoke heat reflection, fire heat reflection, and simple high-temperature objects in real time in an indoor fire situation full of smoke, and creates an early fire suppression environment of a smart building by detecting fire and estimating a fire source. In addition, fire source information is provided continuously even after the initial occurrence of a fire, thereby enabling rapid fire source guidance for firefighters. Effects not mentioned will be clearly understood by those of ordinary skill in the art from the present specification and the accompanying drawings.

1 is a block diagram of an intelligent fire detection device according to an embodiment of the present invention.
2 is a conceptual diagram showing a deep learning model of a deep learning classifier constituting an intelligent fire detection device according to an embodiment of the present invention.
3 is a flowchart of an intelligent fire detection method according to an embodiment of the present invention.
4 is a block diagram of a texture feature extraction unit constituting an intelligent fire detection device according to another embodiment of the present invention.
5 is a flowchart of an intelligent fire detection method according to another embodiment of the present invention.
6 is a flowchart of an intelligent fire detection method according to another embodiment of the present invention.
7 is a conceptual diagram showing a deep learning model of an intelligent fire detection device according to another embodiment of the present invention.

Other advantages and features of the present invention, and a method of achieving them will become apparent with reference to embodiments to be described later in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, and the present invention is only defined by the scope of the claims. Even if not defined, all terms (including technical or scientific terms) used herein have the same meaning as commonly accepted by universal technology in the prior art to which this invention belongs. A general description of known configurations may be omitted so as not to obscure the subject matter of the present invention. In the drawings of the present invention, the same reference numerals are used as much as possible for the same or corresponding configurations. In order to help the understanding of the present invention, some configurations in the drawings may be somewhat exaggerated or reduced.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise", "have" or "have" are intended to designate the existence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification. It is to be understood that the possibility of the presence or addition of other features, or numbers, steps, actions, components, parts, or combinations thereof, or further other features, is not excluded in advance.

The'~ unit' used throughout this specification is a unit that processes at least one function or operation, and may mean, for example, a hardware component such as software, FPGA, or ASIC. However,'~ part' does not mean limited to software or hardware. The'~ unit' may be configured to be in an addressable storage medium, or may be configured to reproduce one or more processors.

As an example,'~ unit' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, procedures, and subs. Routines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The components and functions provided by the'~ units' may be performed separately by a plurality of elements and the'~ units', or may be integrated with other additional elements.

1 is a block diagram of an intelligent fire detection device according to an embodiment of the present invention. Referring to FIG. 1, an intelligent fire detection apparatus 100 according to an embodiment of the present invention includes an infrared thermal image generator 120, a texture feature extractor 140, and a deep learning classifier 160.

The infrared thermal image generator 120 generates infrared thermal image information. The infrared thermal image generator 120 may include an infrared thermal imaging camera (infrared thermal imaging apparatus). In an embodiment, the infrared thermal image generator 120 may collect thermal image information in an infrared long wavelength band. The infrared long wavelength band may be 7.5 to 13 um. The infrared long-wavelength band has the characteristic of penetrating smoke, so it can perform its normal function in a room full of smoke, and it is possible to detect and analyze objects even in dense smoke. In an embodiment, the infrared thermal image generator 120 may provide thermal image information by representing, for example, a temperature of -40°C to 550°C with a 14-bit infrared intensity (a value between -16,384 and -1). .

The infrared thermal image information generated by the infrared thermal image generating unit 120 is transmitted to the texture feature extracting unit 140. The texture feature extraction unit 140 extracts the smoke and fire features from the infrared thermal image information formed with a value between white and black, in order to extract the first statistical texture features and the second statistical texture features from the infrared thermal image information. Extract them.

The texture feature extractor 140 may statistically analyze pixel values of the infrared thermal image information to extract primary statistical texture features. The primary statistical texture features quantify the unique statistical meaning of the pixel values themselves representing the intensity of infrared thermal image information. Deviation), skewness, and kurtosis may include at least two or more. In an embodiment, five primary statistical texture features may be calculated according to Equations 1 to 5 below.

[Equation 1]

[Equation 2]

[Equation 3]

[Equation 4]

[Equation 5]

In Equations 1 to 5,'MNI' is the average value,'VAR' is the variance value,'STD' is the deviation value,'SKE' is the asymmetry,'KUR' is the kurtosis, and I _i,j is the i row and j column The pixel value of the pixel, N _P is the number of pixels, μ is the average value (MNI) of pixel values, and σ is the standard deviation (STD) of pixel values.

The texture feature extractor 140 may quantitatively analyze a correlation between pixels in the infrared thermal image information to extract secondary statistical texture features. The secondary statistical texture features are values that quantify the correlation between the pixel of the thermal image information and the surrounding pixels of the pixel, unlike the primary statistical texture feature that presents the unique information of the pixel in the thermal image information.

In an embodiment, the secondary statistical features are dissimilarity, entropy, contrast, inverse difference, correlation, and similarity between pixels of infrared thermal image information. It may include at least two or more of (Uniformity) and Inverse Difference Moment. In order to extract the secondary texture features, first, a normalized co-occurrence matrix C _i,j is calculated from the infrared thermal image information. The correlation matrix can be calculated according to Equation 6 below.

[Equation 6]

In Equation 6, P _i,j means the frequency of occurrence of gray levels in four directions (up, down, left, right) of the pixels at i and j, and N _G is the pixel of the thermal image information. It means the number of gray levels. In an embodiment, seven secondary statistical texture features may be calculated according to Equations 7 to 13 below based on a correlation matrix.

[Equation 7]

[Equation 8]

[Equation 9]

[Equation 10]

[Equation 11]

[Equation 12]

[Equation 13]

In Equations 7 to 13,'C _ij ' is the correlation matrix,'DIS' is the dissimilarity,'ENT' is the entropy,'CON' is the contrast,'INV' is the inverse difference, and'COR' is the correlation, 'UNI' is the degree of similarity, 'IDM' was reverse and loan Chengdu, 'μ _i', 'μ _j' are each i, the average value of pixel values of the j-th position, 'σ _i', 'σ _j' is i, respectively, It is the standard deviation of the pixel values at the j-th position.

The primary statistical texture features and secondary statistical features extracted by the texture feature extraction unit 140 are input to the deep learning classifier 160. The deep learning classifier 160 analyzes the primary statistical texture features and secondary statistical texture features received from the texture feature extraction unit 140 through a deep learning model to provide a smoke class, a fire class, a smoke heat reflection class, and a fire heat. Probability values are calculated for classes including a reflection class and a simple high-temperature object class.

2 is a conceptual diagram showing a deep learning model of a deep learning classifier constituting an intelligent fire detection device according to an embodiment of the present invention. Referring to FIG. 2, the input layer 182 of the deep learning model 180 includes 5 input nodes 182a consisting of primary statistical texture features and 7 input nodes consisting of secondary statistical texture features ( 182b) may be composed of a total of 12 input nodes. The output layer 186 of the deep learning model 180 may include five output nodes composed of classes. The deep learning model 180 may have four hidden layers 184 between the input layer 182 and the output layer 186.

In an embodiment, the deep learning classifier 160 deep-learns 12 texture features including first-order statistical texture features and second-order statistical texture features, and 1×13 matrix information including ID information of each object. By applying it to the model, probability values for five classes (smoke, fire, heat reflection of smoke, heat reflection of fire, and simple high-temperature objects) can be calculated. Of the four hidden layers, the first hidden layer 184a is composed of 1,024 nodes, and the second, third, and fourth hidden layers 184b, 184c, and 184d are composed of 512, 256, and 64 nodes, respectively.

The deep learning classifier 160 applies 12 or some texture features of an object among 1×13 matrix information as an input node of the input layer 182 of the deep learning model 180 and applies deep learning. Calculate the probability that the object corresponds to five classes, such as smoke, fire, each heat reflection or simple high-temperature object. Deep learning can use a neural network model created through a large amount of existing learning.

The deep learning classifier 160 compares the probability values for each class calculated through the deep learning classification, and analyzes whether the class having the highest probability value among the probabilities of each class is a smoke or fire class. If a result corresponding to the object is smoke or fire, a fire alarm is provided through the network, information on whether the object is smoke or fire, and the location of the object in the image information are provided. If it is determined that the object is not smoke or fire, but is a heat reflection of smoke or fire, the object is not visible on the video screen, but it is determined that a fire has occurred outside the video screen, and a fire alarm is provided, and the location of the object in the video information is provided. .

In general, heat is radiated, and the radiated heat is reflected off objects such as walls and doors. Therefore, when it is recognized as a heat reflection of smoke or fire, it is possible to indirectly know that a fire has occurred, and this is useful information for finding a flower garden. If it is not judged as smoke, fire, or individual heat reflection, and if it is judged as a simple high-temperature object, the fire alarm may not be provided.

3 is a flowchart of an intelligent fire detection method according to an embodiment of the present invention. 3, the infrared thermal image generator 120 generates infrared thermal image information (S10). Infrared thermal image information generated by the infrared thermal image generating unit 120 is transmitted to the texture feature extracting unit 140.

The texture feature extraction unit 140 extracts the smoke and fire features from the infrared thermal image information formed with a value between white and black, in order to extract the first statistical texture features and the second statistical texture features from the infrared thermal image information. They are extracted (S20). The texture feature extraction unit 140 statistically analyzes pixel values of infrared thermal image information to extract primary statistical texture features, and quantitatively analyzes the correlation between pixels in infrared thermal image information to provide a secondary statistical texture. Features can be extracted.

The primary statistical texture features and secondary statistical features extracted by the texture feature extraction unit 140 are input to the deep learning classifier 160. The deep learning classifier 160 analyzes the primary statistical texture features and secondary statistical texture features received from the texture feature extraction unit 140 through a deep learning model to provide a smoke class, a fire class, a smoke heat reflection class, and a fire heat. Probability values for classes including a reflection class and a simple high-temperature object class are calculated (S30).

The deep learning classifier 160 applies 12 texture features including first-order statistical texture features and second-order statistical texture features, and 1×13 matrix information including ID information of each object to the deep learning model. Probability values for five classes (smoke, fire, heat reflection of smoke, heat reflection of fire, and simple high-temperature objects) can be calculated. The deep learning classifier 160 may analyze whether the class having the highest probability value among the probabilities of each class is a smoke or fire class by comparing the probability values for each class calculated through the deep learning classification.

According to an embodiment of the present invention, using deep learning artificial intelligence based on the first statistical texture features and the second statistical texture features based on infrared thermal image information, smoke, fire, each heat reflection, and simple It was experimentally confirmed that high-temperature objects can be detected and identified with high accuracy, and the error rate can be reduced to less than 1% by combining deep learning of the first and second statistical texture features.

According to an embodiment of the present invention, power consumption is low using one sensor (infrared thermal imaging equipment), and it is composed of small sensors and equipment to have small size, light weight, and network characteristics, so that it can be installed anywhere indoors, including smart buildings. It is possible, and by applying the artificial intelligence of deep learning, it is possible to identify smoke, fire, individual heat reflections, and simple high-temperature objects in real time with high accuracy.

In addition, according to an embodiment of the present invention, it is possible to quickly detect a fire by detecting smoke and fire when a fire occurs, and even in the dense smoke of the early and mid-late stages of the fire, smoke, fire, smoke heat reflection, fire heat reflection, simple high-temperature objects In addition to real-time identification, when a fire is visible in an infrared thermal image, the location information of the fire is provided, and when the fire is not visible, the probability of a fire is provided, and the occurrence of non-fire reports can be reduced.

According to an embodiment of the present invention, when an indoor fire occurs, not only detects the fire early, but also continuously provides smoke, fire, and heat reflection information in real time in a situation where the fire continues, so that firefighters quickly reach the fire source. It can be useful for indoor fires in smart buildings, places where early suppression is required in case of fire, and buildings with high military security.

4 is a block diagram of a texture feature extraction unit constituting an intelligent fire detection device according to another embodiment of the present invention. 5 is a flowchart of an intelligent fire detection method according to another embodiment of the present invention. Referring to FIGS. 4 and 5, the texture feature extraction unit 140 provides a primary statistical texture based on environmental information including the material, arrangement structure, temperature and humidity of the fire monitoring area, of the door, wall, and window of the fire monitoring area. Features and secondary statistical texture features may be selected (S110). The texture feature extraction unit 140 may include a primary texture feature selection unit 142, a secondary texture feature selection unit 144, a primary texture feature extraction unit 146, and a secondary texture feature extraction unit 148. I can.

The primary texture feature selection unit 142 is based on the material and arrangement structure of the doors, walls, and windows of the fire monitoring area, among the average values of pixel values, variance values, deviation values, asymmetry and kurtosis, smoke, fire, and smoke heat. One or more primary statistical texture features can be selected in the order of high accuracy for classification of reflection, non-heat reflection, and simple high-temperature objects.

The secondary texture feature selection unit 144 is based on environmental information including temperature and humidity of the fire monitoring area, among the dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and inverse differential inertia between pixels. , Smoke, fire, smoke heat reflection, fire heat reflection, and simple high-temperature objects in the order of high classification accuracy, one or more secondary statistical texture features can be selected.

The information on the primary and secondary statistical texture features selected according to environmental information such as the material and arrangement structure of doors, walls, and windows in the fire monitoring area, temperature and humidity in the fire monitoring area is the primary and secondary texture feature selection section. It may be set in advance at (142, 144).

The primary texture feature extraction unit 146 and the secondary texture feature extraction unit 148 only select primary and secondary statistical texture features selected by the primary texture feature selection section 142 and the secondary texture feature selection section 144. Can be extracted.

The primary texture feature extractor 146 may extract only the primary statistical texture feature selected by the primary texture feature selector 142 from among the five primary statistical texture features from the infrared thermal image information.

The secondary texture feature extractor 148 may extract only the secondary statistical texture feature selected by the secondary texture feature selector 144 from among the seven secondary statistical texture features from the infrared thermal image information.

The deep learning classifier 160 analyzes the primary statistical texture features and secondary statistical texture features received from the texture feature extraction unit 140 through a deep learning model to provide a smoke class, a fire class, a smoke heat reflection class, and a fire heat. Probability values for classes including a reflection class and a simple high-temperature object class are calculated (S130).

According to the embodiment of FIGS. 4 and 5, based on environmental information such as material and arrangement structure of doors, walls, and windows of the fire monitoring area, temperature and humidity of the fire monitoring area, the first statistical texture features and the second statistical By selecting texture features, based on the selected primary and secondary statistical texture features, the accuracy of classification of smoke, fire, smoke heat reflection, fire heat reflection, and simple high-temperature objects is improved, while the computational throughput required for fire prediction. It is possible to predict a fire in real time for many infrared thermal images with limited computational resources. In addition, according to an embodiment of the present invention, since information for analyzing the dynamic aspect of an image acquired by a dynamically moving camera is excluded (excluded) from the secondary statistical texture features and classified for fire prediction, calculation The throughput can be reduced.

6 is a flowchart of an intelligent fire detection method according to another embodiment of the present invention. 7 is a conceptual diagram showing a deep learning model of an intelligent fire detection device according to another embodiment of the present invention. 6 and 7, when an infrared thermal image video image is transmitted from the infrared thermal image generator (infrared thermal imaging device) 120 to the image processing device, the image processing device performs an infrared thermal image by pre-processing the image. High temperature objects (fire, smoke, fire heat reflection, smoke heat reflection, and simple high temperature objects) from the surrounding environment by applying clustering-based image auto-thresholding based on intensity information Objects) are extracted in real time.

The texture feature extraction unit 140 of the image processing apparatus extracts first and second statistical texture features according to Equations 1 to 13 above. The five primary statistical texture features (average value, variance value, deviation value, asymmetry, kurtosis) extracted by the texture feature extraction unit 140 are transmitted to the deep learning classifier 160 of artificial intelligence. It is used in the first stage judgment process of intelligence. The seven secondary statistical texture features (dissimilarity, entropy, contrast, inverse difference, correlation, similarity, inverse difference inertia) extracted by the texture feature extraction unit 140 are the deep learning classifier 160 of artificial intelligence. ), and used in the second-stage judgment process of deep learning artificial intelligence.

In the embodiments of FIGS. 6 and 7, the deep learning model 10 is divided into a two-step class classification process. The first-stage class classification process is an early judgment of simple high-temperature objects, which are non-critical classes. Simple high-temperature objects can be easily determined unlike other classes through the first-stage class classification process. Therefore, it is possible to increase computational efficiency by determining a simple high-temperature object early and reducing the amount of computation for a simple high-temperature object in the entire deep learning model. The second-stage deep learning model is to accurately classify important classes (fire, smoke, fire heat reflection, smoke heat reflection) in real time by considering the first-stage class classification result and the second texture feature.

Deep learning artificial intelligence uses 5 input layers (first statistical texture features), 3 hidden layers and 2 output layers (non-critical class, critical class) based on the first-order statistical texture feature in the first-stage class classification process. The first-stage deep learning model consisting of is applied. In the input layer 12a of the first-stage deep learning model 12, five primary statistical texture features are applied as input values (average value, variance value, deviation value, asymmetry, kurtosis), and the output layer 12b is an important class. And two output values of the non-critical class. The important classes include fire, smoke, fire heat reflection, and smoke heat reflection classes, and the non-critical class may be a simple high-temperature object class. The reason for applying the first-stage deep learning model is to increase computational efficiency by early judging simple hot objects that can be easily judged with a small amount of computation by the first-stage class classification process, reducing the computational amount for simple hot objects in the entire model. It is for sake.

The second-stage deep learning model (14) is based on seven secondary statistical texture features (dissimilarity, entropy, contrast, inverse difference, correlation, similarity, etc.) on the result determined as an important class through the first-stage class classification process. Inverse differential inertia) is added as an input value, and the eight input values are applied as the input layer 14a of the second-stage deep learning model 14. The numbers indicated in parentheses in FIG. 7 indicate the number of nodes in each layer. The second-stage deep learning model 14 is composed of an output layer 14b consisting of 4 hidden layers and 4 output values. At this time, the output value is calculated as a probability value for the important classes (fire, smoke, fire heat reflection, smoke heat reflection). Deep learning artificial intelligence can finally precisely classify four classes that are difficult to classify in real time through a two-stage deep learning model 14. Artificial intelligence calculates the probability value of each object, visualizes the class with the highest probability value for the object, the probability value in infrared image information in real time, and estimates the location of the object in real time.

The intelligent fire detection method according to an embodiment of the present invention can be written as a program that can be executed on a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium. Computer-readable recording media include volatile memories such as SRAM (Static RAM), DRAM (Dynamic RAM), SDRAM (Synchronous DRAM), Read Only Memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Nonvolatile memory such as Electrically Erasable and Programmable ROM (EEPROM), flash memory device, phase-change RAM (PRAM), magnetic RAM (MRAM), resistive RAM (RRAM), ferroelectric RAM (FRAM), floppy disk, hard disk, or The optical reading medium may be, for example, a storage medium such as a CD-ROM or a DVD, but is not limited thereto.

It should be understood that the above embodiments have been presented to aid the understanding of the present invention, and do not limit the scope of the present invention, and various deformable embodiments are also within the scope of the present invention. The technical protection scope of the present invention should be determined by the technical idea of the claims, and the technical protection scope of the present invention is not limited to the literal description of the claims per se, but the invention of the scope of which the technical value is substantially equal. It should be understood that it reaches to.

100: intelligent fire detection device
120: infrared thermal image generation unit
140: texture feature extraction unit
160: deep learning classifier
180: deep learning model
182: input layer
184: hidden layer
186: output layer

Claims (13)

Texture feature extraction that extracts primary statistical texture features by statistically analyzing pixel values from infrared thermal image information, and extracts secondary statistical texture features by quantitatively analyzing the correlation between pixels from the infrared thermal image information part; And
Probability values for classes including a smoke class, a fire class, a smoke heat reflection class, a fire heat reflection class, and a simple high temperature object class by analyzing the first statistical texture features and the second statistical texture features through a deep learning model And a deep learning classifier that calculates
The primary statistical texture features include at least two or more of an average value, a variance value, a deviation value, asymmetry, and kurtosis of the pixel values, and the secondary statistical texture features include dissimilarity, entropy, and contrast between the pixels. An intelligent fire detection device comprising at least two or more of an inverse difference degree, a correlation degree, a similarity degree, and an inverse difference degree of inertia. delete The method of claim 1,
The input layer of the deep learning model includes 12 input nodes consisting of the first statistical texture features and the second statistical texture features, and the output layer of the deep learning model includes 5 output nodes consisting of the classes. And the deep learning model has four hidden layers between the input layer and the output layer. The method of claim 1,
The texture feature extraction unit,
A primary texture feature selection unit for selecting the primary statistical texture features from among the average value, variance value, deviation value, asymmetry and kurtosis of the pixel values based on the material and arrangement structure of the door, wall, and window of the fire monitoring area; And
Based on environmental information including temperature and humidity of the fire monitoring area, the secondary statistical texture features are selected from among the dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and inverse differential inertia between the pixels. Intelligent fire detection device comprising a secondary texture feature selection unit. The method of claim 4,
The texture feature extraction unit,
A primary texture feature extracting unit that extracts only a primary statistical texture feature selected by the primary texture feature selection unit from among the primary statistical texture features from the infrared thermal image information; And
In the infrared thermal image information, the intelligent fire detection apparatus further comprising a secondary texture feature extracting unit for extracting only the secondary statistical texture feature selected by the secondary texture feature selection unit among the secondary statistical texture features. The method of claim 1,
The deep learning model includes a first step deep learning model and a second step deep learning model,
The first input layer of the first-stage deep learning model consists of five input values including the first statistical texture features, and the first output layer of the first-stage deep learning model is a non-critical class corresponding to a simple high-temperature object and Consisting of two output values of an important class excluding the non-critical class, the first-stage deep learning model has at least three hidden layers between the first input layer and the first output layer,
The second input layer of the second-stage deep learning model is composed of eight input values including the important class and seven second-order statistical texture features, and the second output layer of the second-stage deep learning model is fire, smoke, and An intelligent fire detection device comprising four output values corresponding to non-heat reflection and smoke heat reflection classes, and the second-stage deep learning model has at least four hidden layers between the second input layer and the second output layer. Extracting primary statistical texture features by statistically analyzing pixel values from infrared thermal image information;
Extracting secondary statistical texture features by quantitatively analyzing a correlation between pixels from the infrared thermal image information; And
Probability values for classes including a smoke class, a fire class, a smoke heat reflection class, a fire heat reflection class, and a simple high temperature object class by analyzing the first statistical texture features and the second statistical texture features through a deep learning model And calculating them,
The primary statistical texture features include at least two or more of an average value, a variance value, a deviation value, asymmetry, and kurtosis of the pixel values, and the secondary statistical texture features include dissimilarity, entropy, and contrast between the pixels. An intelligent fire detection method comprising at least two of an inverse difference degree, a correlation degree, a similarity degree, and an inverse difference degree of inertia. delete The method of claim 7,
The input layer of the deep learning model includes 12 input nodes consisting of the first statistical texture features and the second statistical texture features, and the output layer of the deep learning model includes 5 output nodes consisting of the classes. Including, wherein the deep learning model intelligent fire detection method having four hidden layers between the input layer and the output layer. The method of claim 7,
Extracting the first-order statistical texture features,
Based on the material and arrangement structure of the door, wall, and window of the fire monitoring area, one or more of the primary statistical texture features including the average value, variance value, deviation value, asymmetry and kurtosis of the pixel values Selecting a; And
And extracting only the one or more primary statistical texture features selected from among the primary statistical texture features from the infrared thermal image information. The method of claim 7,
Extracting the second order statistical texture features,
Among the secondary statistical texture features including dissimilarity, entropy, contrast, inverse difference, correlation, similarity, and inverse differential inertia between the pixels, based on environmental information including temperature and humidity of the fire monitoring area. Selecting one or more secondary statistical texture features; And
And extracting only the one or more secondary statistical texture features selected from among the secondary statistical texture features from the infrared thermal image information. The method of claim 7,
The deep learning model includes a first step deep learning model and a second step deep learning model,
The first input layer of the first-stage deep learning model consists of five input values including the first statistical texture features, and the first output layer of the first-stage deep learning model is a non-critical class corresponding to a simple high-temperature object and Consisting of two output values of an important class excluding the non-critical class, the first-stage deep learning model has at least three hidden layers between the first input layer and the first output layer,
The second input layer of the second-stage deep learning model is composed of eight input values including the important class and seven second-order statistical texture features, and the second output layer of the second-stage deep learning model is fire, smoke, and An intelligent fire detection method comprising four output values corresponding to non-heat reflection and smoke heat reflection classes, and the second-stage deep learning model has at least four hidden layers between the second input layer and the second output layer. A computer-readable recording medium in which a program for executing the intelligent fire detection method of claim 7 or 9 to 12 is recorded.

Images