KR102067994B1 - System for detecting flame of embedded environment using deep learning - Google Patents

Abstract

The present invention relates to a system for detecting a flame of an embedded environment by using deep learning. The system comprises: a detecting unit detecting a flame region according to a chrominance in comparison with the brightness of an input image; a first classification unit inserting the detected flame region as a deep learning input and classifying a region, in which a flame class probability of an output terminal is 75% or more, as a first flame image; a cell separation unit dividing, into cells of N*N units, a region excluding the first flame image classified by the first classification unit (a region in which a flame class probability is less than 75%); and a second classification unit inserting the cells divided into N*N units as a deep learning input, and classifying a region, in which a flame class probability of the output terminal is 50% or more, as a second flame image. According to the present invention, a flame region corresponding to a flame color model can be detected, thereby classifying a flame image through deep learning, and N*N cells of the flame region can be separated to classify a flame shape through deep learning, thereby accurately detecting a flame while maintaining accuracy even in an embedded system with a low-resource environment.

Description

System for detecting flame of embedded environment using deep learning}

The present invention relates to a flame detection system of an embedded environment using deep learning, and more particularly, to detect flame and color at the same time as in the prior art, and to separately perform flame color detection and flame shape detection, thereby reducing low resources ( It is about technology to detect flame accurately while maintaining precision even in embedded system of environment.

The flame detection system is a technology that detects fire early and speeds up suppression and evacuation, and has recently spread the flame detection system to complex buildings to prevent fire early. According to the National Statistical Bureau of Fire Statistics, the damage of fire is expected to be significantly reduced if the flame detection system is further developed and distributed.

On the other hand, the flame detection system up to now has to configure the server, expensive equipment has to be installed or there is a problem of frequent malfunctions. In the existing sensor-based detection system, fire was detected through the installation of smoke and temperature sensors, but there were problems in terms of cost because a large number of sensors had to be installed.

In order to compensate for this, there have been flame detection systems that can detect the flame through image processing using only the image of the camera. Judging from the frequent malfunctions of malfunctions.

Recently, due to the development of deep learning, the malfunction of flame detection using image processing is rapidly reduced, so that almost perfect flame detection can be achieved by image processing alone, but there is a problem that a high specification hardware specification is required.

In addition, the deep learning structure, which is widely used to determine the presence or absence of a flame in the flame detection field, basically requires high performance due to a high hardware specification, a large memory, and a large amount of computation.

Therefore, the present applicant has two light deep learning structures specialized in the color and shape of the flame, and proposes a flame detection system using the deep learning which can be applied to a low specification embedded system while improving the flame detection rate.

Republic of Korea Patent No. 10-1649173 (August 18, 2016.) Republic of Korea Patent No. 10-1716036 (announced on March 13, 2017)

셀로 분리하여 불꽃 모양에 특화된 딥러닝 모델을 통해 50% 이상이 화재로 판단되는 경우 화재영상으로 분류함으로써, 종래에 불꽃 및 색상을 동시에 감지함에 따라 많은 리소스 사용으로 인해 임베디드 시스템 환경에서 불가능했던 불꽃 감지를, 낮은 리소스 환경의 임베디드 시스템에서도 정밀도를 유지하면서 정확하게 감지하는데 있다.The object of the present invention is not to detect the flame and color at the same time as in the prior art, but to detect the flame color model and the corresponding flame region and classify the fire image with a fire image of more than 75% through a deep learning model specialized for the flame color. And the area of the detected fire image

If more than 50% is judged to be a fire through the deep learning model specialized in flame shape by separating into cells, it is classified as a fire image.It detects flame that was impossible in the embedded system environment due to the use of a lot of resources. This is to accurately detect the low-environment embedded system while maintaining the precision.

셀로 분리하여 불꽃 모양 모델을 통해 화재 영상으로 분류함으로써, 낮은 리소스 환경의 임베디드 시스템에서도 정밀도를 유지하면서 정확하게 감지할 수 있다.That is, since the present invention classifies the fire zone only by the fire probability 75% or more after detecting the fire zone through the flame color model, the fire zone detected when the fire probability is less than 75% is determined.

By separating them into cells and classifying them as fire images through a flame model, they can be detected accurately while maintaining precision even in embedded systems in low resource environments.

단위의 셀로 분할하는 셀 분리부; 및

단위로 분할된 셀을 딥러닝 입력으로 넣고, 출력단의 불꽃 클래스 확률이 50% 이상인 영역을 제2 불꽃 영상으로 분류하는 제2 분류부를 포함하는 것을 특징으로 한다.One embodiment of the present invention for achieving the above technical problem is a flame detection system of the embedded environment using deep learning, the detection unit for detecting the flame region according to the luminance difference color difference of the input image; A first classifying unit configured to insert the detected flame region as a deep learning input and classify a region having a flame class probability of at least 75% as the first flame image at an output stage; Areas other than the first flame image classified by the first classification unit (areas with a flame class probability of less than 75%)

A cell separator for dividing the cell into units; And

And a second classifying unit to classify the cell divided into units as a deep learning input and classify a region having a flame class probability of 50% or more as the second flame image.

Preferably, the detector detects the flame region by converting the RGB color model of the input image into the YCbCr color model.

The first classification unit includes a convolution layer for extracting feature data of the detected flame region, a sub sampling layer for removing the noise included in the feature data and reducing the size of the image to confirm the feature data, and the determined feature data. Softmax (NN) that generates a set number of feature vectors from the transformed feature vectors, and classifies regions with a flame class probability of 75% or more as the first fireworks image by generating And a layer.

단위로 분할된 셀의 특징 데이터를 추출하는 Convolution 층과, 특징 데이터에 포함된 노이즈를 제거하고 이미지의 사이즈를 축소하여 특징 데이터를 확정하는 Sub sampling 층과, 확정된 특징 데이터를 일렬로 연결하여 특징 벡터로 변환하는 Fully-connected / Drop out 층, 및 변환된 특징 벡터로부터 설정된 개수의 특징 벡터를 생성하여 불꽃 클래스 확률이 50% 이상인 영역을 제2 불꽃 영상으로 분류하는 Softmax(NN) 층을 포함하는 것을 특징으로 한다.The second classification unit,

The convolution layer extracts feature data of the cell divided into units, the sub sampling layer removing noise included in the feature data and reducing the size of the image to determine the feature data, and connecting the determined feature data in a line. Fully-connected / Drop out layer for converting into a vector, and Softmax (NN) layer for generating a predetermined number of feature vectors from the transformed feature vector to classify a region having a flame class probability of 50% or more as a second flame image It is characterized by.

셀로 분리하여 불꽃 모양에 특화된 딥러닝 모델을 통해 50% 이상이 화재로 판단되는 경우 화재영상으로 분류함으로써, 낮은 리소스 환경의 임베디드 시스템에서도 정밀도를 유지하면서 정확하게 감지할 수 있다.According to the present invention as described above, by detecting a flame region corresponding to the flame color model, through the deep learning model specialized in the flame color to classify an image of fire probability 75% or more as a fire image, and to determine the area of the detected fire image

The deep learning model, which is separated into cells and is specialized in the shape of flame, classifies it as a fire image when more than 50% is judged to be a fire.

셀 분리 과정을 도시한 예시도.
도 11b는 본 발명의 일 실시예에 따른 딥러닝을 이용한 임베디드 환경의 불꽃 감지 시스템의 불꽃 색깔 특화 딥러닝 모델에서 오류를 도시한 예시도.
도 12는 본 발명의 일 실시예에 따른 딥러닝을 이용한 임베디드 환경의 불꽃 감지 시스템의 불꽃 모양 특화 딥러닝 모델을 통한 화재영상 분류를 도시한 예시도.
도 13은 본 발명의 일 실시예에 따른 딥러닝을 이용한 임베디드 환경의 불꽃 감지 시스템에 의해 최종적으로 분류된 불꽃 영상과 불꽃 유사 영상을 도시한 예시도.
도 14a 및 도 14b는 본 발명의 일 실시예에 따른 딥러닝을 이용한 임베디드 환경의 불꽃 감지 시스템의 ImageNet 불꽃 데이터베이스를 도시한 예시도.1 is a flowchart illustrating a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
2A and 2B are deep learning flowcharts of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
3 is an exemplary view illustrating a flame region detection process using a flame color model of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
4A and 4B are exemplary views illustrating flame region detection by a labeling technique of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
5 is an exemplary diagram illustrating a flame image classification process of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
6 is an exemplary view showing a flame color specialized structure by deep learning based on embedded system of a flame detection system of an embedded environment using deep learning according to an embodiment of the present invention.
7A and 7B are exemplary views illustrating a convolution layer of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
8 is an exemplary diagram illustrating a sub sampling layer of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
9A is an exemplary diagram illustrating a Fully-connected / Drop out layer of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
9B is an exemplary diagram illustrating an over fitting phenomenon of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
10 is an exemplary diagram illustrating a Softmax (NN) layer of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
11A illustrates a flame detection system of an embedded environment using deep learning according to an embodiment of the present invention.

Exemplary diagram illustrating a cell separation process.
11B is an exemplary diagram illustrating an error in a flame color-specific deep learning model of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.
FIG. 12 is an exemplary view illustrating fire image classification using a flame shape specialized deep learning model of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention. FIG.
FIG. 13 is an exemplary view illustrating a flame image and a flame-like image finally classified by a flame detection system of an embedded environment using deep learning according to an embodiment of the present invention. FIG.
14A and 14B illustrate an ImageNet flame database of a flame detection system in an embedded environment using deep learning according to an embodiment of the present invention.

Specific features and advantages of the present invention will become more apparent from the following detailed description based on the accompanying drawings. Prior to this, terms or words used in the present specification and claims are defined in the technical spirit of the present invention on the basis of the principle that the inventor can appropriately define the concept of the term in order to explain the invention in the best way. It should be interpreted to mean meanings and concepts. In addition, when it is determined that the detailed description of the known function and the configuration related to the present invention may unnecessarily obscure the subject matter of the present invention, it should be noted that the detailed description is omitted.

단위의 셀(크기가 작은 불꽃 영상)로 분할하는 셀 분리부(300), 및

단위로 분할된 셀을 딥러닝 입력으로 넣고, 출력단의 불꽃 클래스 확률이 50% 이상인 영역을 제2 불꽃 영상으로 분류하는 제2 분류부(400)를 포함하여 구성된다.1, 2A and 2B, the flame detection system S of the embedded environment using deep learning according to an embodiment of the present invention may include a detection unit 100 that detects a flame region according to a luminance contrast color difference of an input image. And the first classifying unit 200 and the first classifying unit 200 for inputting the detected flame region as a deep learning input and classifying a region having a flame class probability of 75% or more as the first flame image. Areas other than the classified first flame image, that is, the area where the flame class probability is less than 75%

A cell separator 300 for dividing a unit cell (a small flame image), and

And a second classifying unit 400 for inserting a cell divided into units as a deep learning input and classifying a region having a flame class probability of 50% or more at the output stage as a second flame image.

Specifically, the detection unit 100 converts an RGB color model of the input image into a YCbCr color model through [Equation 1], and a flame color model (high contrast) that shows a high robustness to a change in illuminance through [Equation 2]. Color difference) to detect the flame region.

[Equation 1]

[Equation 2]

)에 따라 상이하게 검출된 불꽃 영역을 흰색으로 나타낸다.3 is an exemplary view showing a flame region detection process using a flame color model. Here, the left image is the input flame image, and the right image is the threshold value (

The flame region detected differently according to) is shown in white.

는 경험적 임계값(상수)으로, 이 값이 작으면 영상 안의 불꽃을 잘 감지하지만 유사 색상의 다른 개체까지 불꽃 영역으로 검출하는 오류가 발생하게 된다. 따라서, 임계값

는 20 내지 50의 상수로 설정하되, 바람직하게는 실험경과 불꽃 검출률이 가장 높았던 임계값인 35로

를 설정한다.At this time,

Is a heuristic threshold (constant). A small value is a good detection of flames in the image, but an error of detecting other objects of similar color into the flame region occurs. Thus, the threshold

Is set to a constant of 20 to 50, preferably 35 to the threshold value at which the experiment and flame detection rate were the highest.

Set.

값에 따른 화재 검출률 그래프는 아래와 같다.At this time,

The fire detection rate graph according to the value is as follows.

In addition, the flame area detection of the detection unit 100 uses the labeling technique of assigning a unique number to all the pixels belonging to the same object, as shown in FIGS. 4A and 4B. Can be detected. At this time, the detection unit 100 selects the labeling area having the most fire area.

Meanwhile, the first classification unit 200 inserts the fire region detected through the flame color model as an input of the flame color specialized deep learning model, and classifies the fire image as a fire image only when the fire class probability of the output terminal is 75% or more. In this case, as a result of performing the fire zone detection through the flame color model in the image of the color similar to the fire, the case of the color similar to the flame may also be detected.

When the fire-like image is input to the flame-colored deep learning model, the fire class probability of the output stage is mostly between 50% and 75%, which is less than 75% of the flame color-specific deep learning model. If it is classified, it will not be judged by the fire image.

Flame image classification process using the deep learning of the first classification unit 200 is performed by a convolution layer, a sub sampling layer, a fully-connected / drop out layer, and a Softmax (NN) layer, as shown in FIG. The flame color specialized structure by deep learning based on the embedded system according to an embodiment of the present invention is shown in FIG. 6 and Table 1 below.

Referring to FIG. 6 and [Table 1], the convolution operation is performed eight times, and the filter is a 3 × 3 random value filter and the number of filters is 16, 32, 64, 64, 128, 128, 128 from No. 1, respectively. , 128.

Subsequently, the max sub sampling operation is performed once in No. 2, 4, 7, 10, and 13 after the convolution operation.

After the operation up to No. 13, the fully connected operation No. 14 performs the function of connecting all the extracted features to generate the final feature vector, and extracts a total of 3200 feature vectors by simply concatenating the feature vectors.

No. 15 reduces the over fitting of deep learning by making the value of some feature vectors to 0 through drop out from a total of 3,200 feature vectors.

In the last No. 16, 3,200 feature vectors, which have been advanced to drop out, are put into the neural network input and set to the number of classes that learn the number of nodes in the neural network output layer.

In the case of flame detection, the neural network output layer node is set to two because the class to be learned is two flame / non-flame.

Hereinafter, the detailed procedure of the first classification unit 200 according to an embodiment of the present invention is as follows.

First, as illustrated in FIGS. 7A and 7B, the first classifying unit 200 extracts feature data of a flame region detected through the convolution layer.

Second, the first classifier 200 determines the feature data by removing noise included in the feature data through the sub sampling layer and reducing the size of the image as shown in FIG. 8 (extracted through the convolution layer). A process of reducing the size of the feature image by selecting only the optimal features among the feature images).

Third, the first classifying unit 200 converts the feature data determined through the Fully-connected / Drop out layer in a line and converts the feature data into a feature vector as shown in FIG. 9A.

The Fully-connected layer performs convolution and sub-sampling by concatenating the final feature data extracted in a row, converting them into feature vectors, and the Drop Out layer makes some feature vectors 0 to deep learning. Reduce over fitting

Here, as shown in FIG. 9B, the over fitting refers to a phenomenon in which a large amount of layers or excessive learning are allowed, but the error is excessively classified as the above cause.

That is, after the convolution layer and the sub-sampling layer of 13 steps, all extraction features are connected to generate a final feature vector. In this case, No. 14 of Table 1 simply extracts a total of 3200 feature vectors by concatenating the feature vectors, and No. 15 of Table 1 selects some feature vectors through drop out from a total of 3200 feature vectors. Set to 0 to reduce the over fitting of deep learning.

Fourth, as shown in FIG. 10, the first classifier 200 generates a predetermined number of feature vectors from the feature vectors transformed through the Softmax (NN) layer, thereby generating a region having a flame class probability of at least 75%. Classify as fireworks image.

In addition, the characteristic filter coefficients of the convolution layer and the weights of the NN forward propagated to the Softmax layer are learned through a backpropagation algorithm. At this time, the back-propagation algorithm corrects the weight by inputting the error between the target value and the output value in the reverse direction of the structure, and decreases the error as the learning proceeds to output the desired result.

단위의 셀(크기가 작은 불꽃 영상)로 분할한다. 이때, 불꽃의 색깔이 다소 부족한 화재 영상은 화재 영상으로 분류되지 않을 수 있다.Meanwhile, the cell separator 300 may determine an area having a flame class probability of less than 75%, which is an area other than the first flame image classified by the first classification unit 200.

The cell is divided into units (small flame images). At this time, the fire image that the color of the flame is somewhat insufficient may not be classified as a fire image.

값에 따라 딥러닝 구조에 입력되었을 때 결과에 많은 영향을 미치게 된다. 실험 결과

이상일 경우에, 너무 작은 셀로 분할되기 때문에 딥러닝 학습이 되지 않을 우려가 이다.

일 경우, 최소한의 불꽃으로 판단될 수 있는 특징들이 남아있어 학습이 가능한바, 본 발명의 일 실시예에서는

로 결정하였다.Also, divided

Depending on the value, it will have a big impact on the result when it is entered into the deep learning structure. Experiment result

In case of abnormality, deep learning is not possible because it is divided into too small cells.

In this case, the features that can be determined to be the minimum flame remains to be learned, in one embodiment of the present invention

Determined.

로 분할된 셀들은 딥러닝 구조에 맞추어 일정한 크기로 조정하게 된다. like this

The cells divided by are adjusted to a predetermined size according to the deep learning structure.

128의 이미지가 들어가므로

로 분할된 셀들은 모두 128

128의 일정한 크기로 조정된다.The input of the first convolution operation of the flame-shaped deep learning structure according to an embodiment of the present invention is 128

As 128 images enter

The cells divided by are all 128

It is scaled to 128 constant sizes.

셀 분리 과정을 거치는 이유는 다음과 같다.At this time, before entering into the flame-shaped deep learning structure

The reason for the cell separation process is as follows.

The whole image and the cell-separated image have their own characteristics. The entire image can be classified into an image specialized in the color information of the flame, and through the learning, a deep learning structure specialized in the flame color information can be formed. The cell-separated image may be classified into an image specialized in flame shape information, and a deep learning structure specialized in flame shape information may be formed through learning.

셀 분리 과정을 거쳐 불꽃 모양 특화 딥러닝 구조로 수행하게 되면 불꽃 영상으로 분류될 수 있다.In addition, a flame image that is not classified as a flame image due to a lack of flame color in a flame color specialized deep learning structure is also included.

If the cell separation process is performed in a flame-shaped deep learning structure, it may be classified as a flame image.

셀 분리 과정을 도시한 도면이다. 우측 영상에서 우상단 셀은 불꽃 색깔 모델을 통해 검출된 불꽃 영역에서 불필요한 영역으로 볼 수 있다.11A illustrates a flame image not classified as a flame image.

A diagram illustrating a cell separation process. In the right image, the upper right cell can be seen as an unnecessary region in the flame region detected by the flame color model.

In addition, as shown in FIG. 11B, fire-like images such as fluorescent lamps and flashes have a lot of flame colors distributed when looking at the image due to light bleeding and the like, thereby causing an error in a flame color-specific deep learning model.

When the entire image is separated into cells, the cells in the upper right and the lower left are not flames, but the shape of the fluorescent lamp is clear, so it is not judged as a fire.

단위로 분할된 셀을 딥러닝 입력으로 넣고, 출력단의 불꽃 클래스 확률이 50% 이상인 영역을 제2 불꽃 영상으로 분류한다.On the other hand, the second classification unit 400

A cell divided into units is input to a deep learning input, and a region having a flame class probability of 50% or more at the output terminal is classified as a second flame image.

단위로 분할된 셀의 특징 데이터를 추출한다.First, the second classification unit 400 through the convolution layer

Feature data of a cell divided into units is extracted.

Second, the second classifier 400 determines the feature data by removing noise included in the feature data through the sub sampling layer and reducing the size of the image.

Third, the second classifier 400 converts the feature data determined through the Fully-connected / Drop out layer into a feature vector.

Fourth, the second classifier 400 generates a set number of feature vectors from the feature vectors transformed through the Softmax (NN) layer to classify the region having the flame class probability of 50% or more as the second fireworks image.

셀 분리 이미지들은 불꽃의 모양에 특화된 이미지로 분류된다.Specifically, the detected flame zone

Cell-separated images are classified into images specialized for the shape of the flame.

Hereinafter, the fire image classification through the flame shape specialized deep learning model shown in FIG. 12 and [Table 2] is as follows.

First, the convolution layer operation is performed five times, and the filter is a 3 × 3 random value filter and the number of filters is 8, 16, 32, 64, 128 from No. 1 in [Table 2]. This convolution operation yields a fast fire detection time.

Convolution layer operation Next, perform sub sampling operation once in No. 2, 4, 6, 8, 10 of [Table 2].

Subsequently, after the calculation up to No. 10, the fully connected operation, No. 11 in [Table 2], serves to connect all the extracted features to generate the final feature vector, and extracts a total of 2,048 feature vectors by simply concatenating the feature vectors. do.

In addition, No.12 in [Table 2] reduces the over fitting of deep learning by making the value of some feature vectors to 0 through drop out from a total of 2,048 feature vectors.

Finally, in No. 13 of [Table 2], 2,048 feature vectors that have been advanced to drop out are put as neural network inputs and set as the number of classes for learning the number of nodes in the neural network output layer. At this time, the finally sorted flame image and the flame-like image is the same as the example shown in FIG.

In the case of flame detection, the neural network output layer node is set to two because the class to be learned is two flame / non-flame.

Hereinafter, an experimental result for verifying a flame detection system S of an embedded environment using deep learning according to an embodiment of the present invention will be described below.

The embedded board used for the experiment was a Raspberry pi-3, and the hardware specification consists of a Quad Core 1.4GHz CPU, 1GB of RAM, and a Broadcom VideoCore IV MP2 400MHz GPU. The development tools in the Raspbian operating system used python3 and the Tensorflow library.

  In addition, the reliability of the flame database of ImageNet was evaluated by experimenting with the existing deep learning structure, VGG and Inception-v3 structure. The ImageNet database consists of a total of 5,063 images of fireworks, fluorescent lights, and flashes, of which 3,545 (70%) were used for learning and 1,518 (30%) were used for the experiment. The ImageNet Flame Database includes several flame images such as bonfire, stove, firewood, and the like as shown in FIGS. 14A and 14B.

[Table 3] is a table comparing the resource occupancy according to the flame detection system of the embedded environment using the deep learning according to an embodiment of the present invention, [Table 4] is using the deep learning according to an embodiment of the present invention The table compares the flame detection rate of the flame detection system in the embedded environment and the conventional deep learning structure.

In the deep learning structure, a detection rate of 98.2% was measured, and a detection rate of 98.7% for a VGG structure and 99.6% for an Inception-v3 structure were measured. Compared with the conventional deep learning structure, the average result was 0.95% lower, but this result is from the lighter configuration than the existing deep learning structure in order to develop a deep learning structure suitable for an embedded system.

In addition, light layering for the development of deep learning architectures for embedded systems requires low resource occupancy. However, if a light layer is formed, a problem of low fire detection rate occurs. Therefore, it is important to find an optimal trade-off point between resource occupancy and fire detection rate, and it is possible to establish a resource occupancy of 60% or less when applied to an embedded fire detection system.

In an embodiment of the present invention, as shown in Table 5, the resource occupancy rate of 46.5% and the fire detection rate of 98.2% were confirmed.

In addition, Table 6 shows the flame detection time of the deep learning structure and the conventional deep learning structure according to an embodiment of the present invention. In the deep learning structure according to an embodiment of the present invention 19.3 seconds were measured, 23.4 seconds in the VGG structure, 31.2 seconds in the Inception-v3 structure is measured, flame detection on average 8 seconds faster than the conventional deep learning structure The time is shown.

TABLE 6

In summary, according to the deep learning structure according to an embodiment of the present invention, the resource occupancy is reduced by 29.86% on average compared to the existing deep learning structure. That is, the proposed deep learning structure is designed to be lighter than the existing deep learning structure in order to optimize the embedded system, thus using less resources.

In addition, the fire detection rate is 0.95% lower than the conventional deep learning structure. Since the layers are lightly layered to develop a deep learning structure suitable for an embedded system, a problem of low fire detection rate occurs. When applied to the actual fire detection embedded system can be built up to secure a resource share of less than 60%, in one embodiment of the present invention confirmed a 46.5% resource share and a fire detection rate of 98.2%.

In addition, according to an embodiment of the present invention, the fire detection time is reduced by an average of 8 seconds compared to the existing deep learning structure. In other words, the proposed deep learning structure is designed with a much smaller layer than the existing deep learning structure, so the time required for deriving the recognition result is quick.

As described above and described with reference to a preferred embodiment for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as shown and described as such, it is a deviation from the scope of the technical idea It will be understood by those skilled in the art that many changes and modifications can be made to the invention without departing from the scope of the invention. Accordingly, all such suitable changes, modifications, and equivalents should be considered to be within the scope of the present invention.

S: Flame Detection System in Embedded Environments Using Deep Learning
100: detector
200: first classification unit
300: cell separator
400: second classification unit

Claims (4)

A detector for detecting a flame region according to a luminance difference color difference of the input image;
A first classifying unit configured to insert the detected flame region as a deep learning input and classify a region having a flame class probability of at least 75% as the first flame image at an output stage;
An area other than the first flame image classified by the first classification unit (an area having a flame class probability of less than 75%)

A cell separator for dividing the cell into units; And

A second classifying unit configured to classify a cell divided into units as a deep learning input and classify a region having a flame class probability of 50% or more as a second flame image at an output terminal;
The second classification unit,

A convolution layer extracting feature data of cells divided into units, a sub sampling layer removing noise included in the feature data and reducing the size of the image to determine the feature data, and connecting the determined feature data in a line Fully-connected / Drop out layer to convert to vector, and Softmax (NN) layer that generates a set number of feature vectors from the transformed feature vector to classify the region having a flame class probability of 50% or more as a second flame image
Flame detection system of an embedded environment using deep learning, comprising.
The method of claim 1,
The detection unit,
Flame detection system of the embedded environment using a deep learning characterized in that for detecting the flame region by converting the RGB color model of the input image to a YCbCr color model.
The method of claim 1,
The first classification unit,
A convolution layer for extracting feature data of the detected flame region;
A sub sampling layer that removes noise included in the feature data and reduces the size of the image to determine the feature data;
Fully-connected / Drop out layer that connects the determined feature data in series and converts them into feature vectors, and
Softmax (NN) layer that generates a predetermined number of feature vectors from the transformed feature vectors and classifies the region having the flame class probability of 75% or more as the first flame image.
Flame detection system of an embedded environment using deep learning, comprising.
delete

Images