KR20220071143A - Lightweight FIRE-DET flame detection method and system - Google Patents

Abstract

The present invention discloses a light-weight FIRE-DET flame detection method and system. First, a data set including a complex environment is constructed; Next, build FIRE-DET, a single-step detection model, and use a FIRE-Net network in which a multi-convolutional combination structure is stacked on the backbone of the model, while reducing the number of convolutional channels; Using BiFPN as the model feature to fuse networks to improve the effect of multi-scale feature fusion; augmenting improved spatial attention mechanisms to enhance flame characteristics; Prediction and regression are performed by inputting the extracted features into the Class/Box Net layer; Finally, by training the FIRE-DET model using the data set, we obtain the detector used to detect the flame. The present invention has an identification rate of 97.55% and a flame detection speed of 45 frames/s, which can be used for flame detection and early warning, and has relatively good robustness and wide application value.

Description

Lightweight FIRE-DET flame detection method and system

The present invention belongs to the field of image processing and fire prevention technology, and more particularly, to a light-weight FIRE-DET flame detection method and system.

In recent years, as the research of deep learning progresses more and more in-depth, the field of application is getting wider. The conventional flame detection system is composed of a researcher combining computer vision, and there are mainly the following types. (1) Image segmentation method based on color space: submitted by researchers such as Nurul SB, using image augmentation technology, RGB and YCbCr color models, and separating fire pixels from the background under predetermined conditions, It is a method to detect fire by separating and contrasting luminance and chromaticity from Researchers such as Dmytro Peleshko have presented a method based on fusing color segmentation and moving object detection, which is superior to other algorithms with maximum resource limitations. Researchers such as Teng Wang and others have built several expert systems that include color dispersion, similarity, and center of mass motion to identify flames. (2) Method based on image slope and integration: Researchers such as OUYANG Ji-neng submitted edge slope of flame image for study, and selected the edge slope of color B as the basis for judging flame and interference images. took A flame identification model based on the image edge slope was constructed by fitting the curve of the image edge slope of the mass sample. Alexander F submitted a method that detects flames by calculating image color and shape features and integral calculations, but does not increase the processing time rapidly. (3) Detection method based on infrared image: Kewei Wang replaces the conventional hand-experience method with a 9-layer convolutional neural network IRCNN to extract infrared image features, and then uses the extracted features to perform a linear support vector machine (support vector machine) ) to detect fire. (4) Image augmentation and color space-based method: Researchers such as Nurul SB use image augmentation technology, RGB and YCbCr color models to separate fire pixels from the background under predetermined conditions, and obtain luminance and chromaticity from the original image. A method for detecting fire by isolating and contrasting was presented. (5) Detection method based on the fusion of segmentation and classification: Researchers such as Andrew J. D have submitted a flame detection method based on super Pickbell segmentation, which uses the technology of super pixel segmentation to scan drawings. After dividing, the flame is detected using the classification method again. Researchers such as Naigong Yu and others presented a method of extracting a region suspected of being a flame from a video using a method combining motion feature detection and color feature detection, and then classified the extracted suspect region using a convolutional neural network of two streams. method was presented. (6) Detection method based on deep learning: It is a method to detect and visualize flames using a neural network detection method. For example, researchers from Huiten et al. submitted a flame detection method based on the Faster-RCNN model; Researchers such as Ducheon City (杜晨錫) submitted a video flame detection method based on YOLOv2. Researchers such as Khan Muhammad have presented a novel, energy-saving, computationally efficient CNN architecture.

Methods based on deep learning are superior to other methods in terms of generalization, but training of deep convolutional neural networks has relatively high demands on the data set capacity and computer usage. Aiming at this problem, this patent submits a light-weight FIRE-DET flame detection method and system.

The present invention was created to solve the problems existing in the prior art, and the purpose is to detect a flame in real time and maintain a relatively good accuracy, and a light-weight FIRE-DET flame detection method that provides a visualization effect and to provide a system.

The light-weight FIRE-DET flame detection method is detailed,

(1) build a FIRE-DET model; the FIRE-DET model includes a feature extraction network, a feature fusion network, an image segmentation network and a predictive identification network; The feature extraction network is constructed by stacking multi-convolution combination structures, and reduces the number of convolutional channels to lower the parameter amount of the detection model; the feature fusion network is a BiFPN network; The image segmentation network performs deconvolution and convolution manipulation on the features after fusion, then obtains an attention drawing and fuses them with the features obtained by the feature fusion network;

(2) forming a data set by performing pre-processing and normalization on a video frame image including a pre-obtained complex environment; In addition, training the FIRE-DET model to obtain a model M used to detect the flame;

(3) using the model M to predict the frame and type of the flame object, and then obtain and preserve the location information of the flame; Visualize the preserved flame target position in the original moving image; After the ratio of the sum of the target area of each flame in the video frame and the area of the original video frame reaches a predetermined threshold, sending a flame alarm; includes.

Further, the operation process of the feature extraction network described in step (1) is as follows,

Normalizes the input drawing to a size of 416×416 to generate IMG0; IMG0 after normalization is taken as an input of the multi-convolutional combination network and calculated to obtain F1; After performing a maximum pooling operation with a convolution kernel of 2×2 on F1, the obtained Pool1 is used as an input to the multi-convolutional combination network and calculated to obtain F2; After performing the maximum pooling operation with the convolutional kernel of 2×2 on F2, the obtained Pool2 is used as an input to the multi-convolutional combination network and calculated to obtain F3; After performing the maximum pooling operation with a convolutional kernel of 2×2 on F3, the obtained Pool3 is used as an input to the multi-convolutional combination network and calculated to obtain F4; After performing the maximum pooling operation with a convolutional kernel of 2×2 on F4, the obtained Pool4 is used as an input to the multi-convolutional combination network and calculated to obtain F5; After performing a maximum pooling operation with a convolutional kernel of 2×2 on F5, a convolution operation with a size of a convolutional kernel of 3×3 is performed to obtain F6; After performing a maximum pooling operation with a convolution kernel of 2×2 on F6, a convolution operation with a size of a convolution kernel of 3×3 is performed to obtain F7.

Furthermore, the operation process of the feature fusion network described in step (1) is as follows,

The features obtained through the multi-convolutional combination structure of the feature extraction network are also obtained using F3, F4, F5, F6 and F7 as inputs of the BiFPN network to converge the features, and then outputs C1, C2, C3, C4 and C5 are obtained;

Taking C1, C2, C3, C4 and C5 as inputs of the BiFPN network, fusing the features again to obtain outputs D1, D2, D3, D4 and D5; The process of the BiFPN network is as follows,

1) In the network, there are 5 inputs denoted as Input1, Input2, Input3, Input4 and Input5, respectively;

2) After performing deconvolution operation of 2×2 on Input1, arrange and operate with Input2 to obtain A1; After performing a 2×2 deconvolution operation on A1, arrange and operate with Input3 to obtain A2; After performing a 2×2 deconvolution operation on A2, arrange and operate with Input2 to obtain A3;

3) After performing 2×2 deconvolution operation on A3, arrange with Input5 and operate to obtain B5; After performing the maximum pulling operation of 2×2 on B5, arrange and operate with Input4 and A3 to obtain B4; After performing the maximum pulling operation of 2×2 on B4, arrange and operate with Input3 and A2 to obtain B3; After performing the maximum pulling operation of 2×2 on B3, arrange and operate with Input2 and A1 to obtain B2; After performing the maximum pulling operation of 2×2 on B2, arrange and operate with Input1 to obtain B1;

4) B1, B2, B3, B4, and B5 in 1) to 3) are taken as outputs of the feature fusion network.

Furthermore, the implementation process of step (2) is as follows,

(21) Extract the video frame set from the video by applying the method of adopting a key frame to the published flame video, and mark the location of the flame using the image marking tool on the video frame set and build a tag data set and;

(22) Make the binary image corresponding to each moving image among the training data sets follow the tag data set, and set the position of the flame indicated by the original drawing corresponding to the binary image to 1. , other parts are set to 0 to form a tag data set of FIRE-DET's rational image, and finally, a data set of FIRE-DET is built on a video frame set, a flame target tag data set and a tag data set of a logical image, ;

(23) The target pixel value based on the normal distribution of the moving picture frame image among the moving picture frame sets is increased, and the data of the data set is augmented by randomly generating horizontal mirroring for the moving picture frame image.

Further, the loss function of the image segmentation network described in step (3) is

이며,

is,

where x is the attention diagram output by the image segmentation network, z is the rational image representing the flame target; The three attention diagrams output by the image segmentation network are each subjected to a loss function and a feedback calculation using a rational image to represent the flame.

The present invention is

an image preprocessing module used to read a moving picture frame image and perform preprocessing and normalization on the moving picture frame image;

a flame detection module used to obtain location information of a flame target by detecting a video frame image after normalization using the trained FIRE-DET model;

a flame area visualization module used to realize the visualization of flame tracking by marking the corresponding flame target in the original video image based on the flame target position preserved by the flame detection module;

A flame alarm module used to continuously monitor the video and send a flame alarm to notify the user when the flame target exceeds the predetermined area in the video frame; Provides a lightweight FIRE-DET flame detection system comprising: do.

The present invention includes a memory, a processor, and a computer program that is stored in the memory and can be operated in the processor, and when the computer program is loaded to the processor, the light-weight class FIRE-DET flame detection method is implemented. FIRE-DET flame detection system is further provided.

In contrast to the prior art, the present invention builds a lightweight FIRE-DET flame detection model and uses a multi-convolution kernel combination structure in the FIRE-DET model to reduce the channel amount of the feature extraction network layer; Introduce the BiFPN network and fuse the features obtained from the backbone network to obtain a feature diagram with multi-scale information and improve the detection accuracy of the model; At the same time, by introducing an image segmentation module and performing deconvolution and convolution operation on the feature diagram obtained by the feature fusion network module, the attention diagram is obtained and used for model detection, thereby improving the detection accuracy of the model; In addition, it improves the convergence speed of the model by calculating the loss functions of the three attention drawings of the image segmentation module, respectively; The present invention has an identification rate of 97.55% and a flame detection speed of 45 frames/s, which can be used for real-time flame detection and early warning, and has relatively excellent robustness and a wide range of application values.

1 is a structural explanatory diagram of a FIRE-DET model in the present invention;
2 is a detection flowchart of the present invention;
3 is a backbone FIRE-NET structure diagram of a FIRE-DET model in an embodiment of the present invention;
4 is a BiFPN structural diagram of a FIRE-DET model in an embodiment of the present invention;
5 is a structural diagram of a multi-convolution combination according to an embodiment of the present invention;
6 is a layer structure diagram of a FIRE-DET model in an embodiment of the present invention;
7 is a flowchart of fusing a feature diagram and an attention diagram in a FIRE-DET model according to an embodiment of the present invention;
8 is a process diagram obtained during detection in an experiment of an embodiment of the present invention; Here, (a) is a video frame after pre-processing; (b) is an attention diagram obtained after going through the image segmentation network among the FIRE-DET models; (c) is a binary diagram of the flame positions obtained in the notation (a);
9 is a flowchart of experimental detection according to an embodiment of the present invention.

Hereinafter, the present invention will be described in further detail in conjunction with the drawings.

In this embodiment, a large number of variables are involved, and each variable is currently described as shown in Table 1 below.

Table 1 Variable Description Table

This implementation method uses the flame video data of the actual scene, and the video includes flames of different colors, flames of different shapes, flames with small targets, special flames, and lights close to the color of the flame. Treatment, detection of flame location, and visualization.

The light-weight FIRE-DET flame detection method provided by the present invention includes the following steps in detail.

Step 1: Construct a FIRE-DET model as shown in FIG. 1 .

The FIRE-DET model includes a feature extraction network, a feature fusion network, an image segmentation network and a predictive identification network; The feature extraction network is constructed by stacking multi-convolution combination structures, and reduces the number of convolutional channels to lower the parameter amount of the detection model; the feature fusion network is configured as a BiFPN network to enhance the effect of multi-scale feature fusion; The image segmentation network performs deconvolution and convolution manipulation on the features after fusion, then obtains an attention diagram and fuses them with the features obtained by the feature fusion network.

1. The feature extraction network uses a FIRE-Net network in which a multi-convolutional combination structure is stacked and reduces the number of convolutional channels.

1) The input performs the convolution operation with a convolution kernel of 1×1 three times, and records the results of the 1st and 3rd convolutions as F11 and F12.

2) Three convolutions are performed on the input. First, the convolutional kernel is 3×3, then the convolutional kernel is 1×1, and finally, the convolutional kernel is After the 3×3 convolution operation, the results of the 1st and 3rd convolutions are recorded as F21 and F22.

3) Three convolutions are performed on the input. First, the convolutional kernel is 9×9, then the convolutional kernel is 1×1, and finally, the convolutional kernel is After the 9×9 convolution operation, the results of the first and third convolutions are recorded as F31 and F32.

4) F11, F12, F21, F22, F31 and F32 in each of 1) to 3) are fused by weight, and F is output.

The feature extraction network structure is as shown in FIG. 3 , and the layer structure shown in FIG. 6 is constructed, and the operation process is as follows.

(11) generate IMG0 by normalizing the input drawing to a size of 416×416;

(12) Calculate F1 by taking the normalized IMG0 as an input to the multi-convolutional combinatorial network;

(13) After performing a maximum pooling operation with a convolutional kernel of 2×2 on F1, the obtained Pool1 is used as an input to the multi-convolutional combination network and calculated to obtain F2;

(14) F3 is obtained by calculating Pool2 obtained after performing the maximum pooling operation with a convolutional kernel of 2×2 on F2 as an input to the multi-convolutional combination network;

(15) F4 is obtained by calculating Pool3 obtained after performing the maximum pooling operation with a convolutional kernel of 2x2 on F3 as an input to the multi-convolutional combination network;

(16) F5 is obtained by calculating Pool4 obtained after performing the maximum pooling operation with a convolutional kernel of 2×2 on F4 as an input to the multi-convolutional combination network;

(17) After performing the maximum pooling operation with the convolution kernel of 2×2 on F5, the convolution operation with the size of the convolution kernel of 3×3 is performed to obtain F6;

(18) After performing the maximum pooling operation with the convolution kernel of 2×2 on F6, the convolution operation with the size of the convolution kernel of 3×3 is performed to obtain F7.

2. BiFPN network combination builds a feature fusion network of the model to enhance the effect of multi-scale feature fusion.

As shown in FIG. 4, the detailed feature fusion process of the BiFPN network is as follows.

1) There are 5 inputs marked as Input1, Input2, Input3, Input4 and Input5 respectively in the network.

2) After performing deconvolution operation of 2×2 on Input1, arrange and operate with Input2 to obtain A1; After performing a 2×2 deconvolution operation on A1, arrange and operate with Input3 to obtain A2; After performing a 2×2 deconvolution operation on A2, arrange and operate with Input2 to obtain A3.

3) After performing 2×2 deconvolution operation on A3, arrange with Input5 and operate to obtain B5; After performing the maximum pulling operation of 2×2 on B5, arrange and operate with Input4 and A3 to obtain B4; After performing the maximum pulling operation of 2×2 on B4, arrange and operate with Input3 and A2 to obtain B3; After performing the maximum pulling operation of 2×2 on B3, arrange and operate with Input2 and A1 to obtain B2; After performing the maximum pulling operation of 2×2 on B2, arrange and operate with Input1 to obtain B1.

4) B1, B2, B3, B4, and B5 in each of 1) to 3) are taken as outputs of the BiFPN network.

The operation process of the feature convergence network is as follows.

(21) using F3, F4, F5, F6 and F7 as inputs of the BiFPN network to fuse features to obtain C1, C2, C3, C4 and C5;

(22) Take C1, C2, C3, C4 and C5 as inputs of the BiFPN network, and then fuse the features again to obtain D1, D2, D3, D4 and D5.

3. The FIRE-DET model image segmentation network is,

(31) perform a convolution operation with a convolution kernel of 3×3 on C5 to obtain AttentionMap1;

(32) perform a convolution operation with a convolution kernel of 3×3 on C4 to obtain AttentionMap2;

(33) Performing a convolution operation with a convolution kernel of 3×3 on C3 to obtain AttentionMap3.

As shown in FIG. 8, (a) is a video frame after preprocessing, (b) is an attention diagram, (c) is a binary image, and the network pays more attention to the flame position among (a) In order to improve the function, a loss function is defined between the attention diagram obtained after deconvolution and convolution in the feature diagram and the rational image of the flame target as a loss function, and the loss is calculated.

에 의해 계산되며,The loss function is

is calculated by

Here, x is the attention diagram output by the image segmentation network, and z is the rational image representing the flame target. The three attention diagrams output by the image segmentation network can each use the loss function to perform a rational image representing a flame and a feedback calculation, and to quickly lower the loss function of the network to reach convergence.

4. The FIRE-DET model prediction identification network includes the following as shown in FIG. 7 .

(41) After duplicating and assembling AttentionMap1 so that the number and size of channels in D5 match, calculate the dot product with it, and then perform convolution operation with the size of the convolution kernel of 3×3 to obtain Map1.

(42) After replicating and assembling AttentionMap2 so that the number and size of channels in D4 match, calculate the dot product with it, and then perform convolution operation with the size of the convolution kernel of 3×3 to obtain Map2.

(43) After duplicating and assembling AttentionMap3 so that the number and size of channels in D3 match, calculate the dot product with it, and then perform convolution operation with the size of the convolution kernel of 3×3 to obtain Map3.

(44) Map1, Map2, and Map3 are the classification probability of outputting the prediction candidate window among the video frames to which the model corresponds, and the coordinate deflection amount relative to the corresponding standard size window. get information

Step 2: As shown in Fig. 2, after preprocessing the actual flame video data including the pre-obtained complex environment, training the FIRE-DET model to obtain M, specifically including the following.

(1) By applying the frame extraction method to the actual flame video data, one frame is cut out every 30 frames from the video to construct the training data set P1, where P1 is {Frame _1, Frame2, … , Frame _N }, Frame _N is the cut Nth video frame, and the training data set is marked with the location of the flame with the labelImg image marking tool to construct the tag data set L1, where L1 is {Label _1, Label2, … , Label _N } and Label _N is the position of the flame in Frame _N , but each Label is (x ₁ ,y ₁ ,x ₂ ,y ₂ ), among which (x ₁ ,y ₁ ) is the position of the upper left corner of the flame and (x ₂ ,y ₂ ) represents the position of the lower right angle of the flame. For the logical image corresponding to each video frame image in the training data set P1, the flame position of the logical image is set to 1 based on the tag data set L1, and the other parts are set to 0 to form the tag data set L2 of FIRE-DET. Finally, P1, L1 and L2 constitute the data set Data of FIRE-DET. The pixel value based on the normal distribution of the moving picture frame image among the moving picture frame sets is increased, and the data of the data set is augmented by randomly generating horizontal mirroring for the moving picture frame image.

(2) Set the pre-trained weight to a random value and set the input dimension of the FIRE-DET model to (416, 416, 3).

(3) FIRE-DET model parameters, including applying the Adam gradient descent method, setting the learning rate to 1×10-4, setting the number of sampling runs to 4, and setting the loss function as a cross entropy function to set

(4) Using P1 in Data1 as input values and L1 and L2 as monitoring values, train among FIRE-DET models to obtain model M.

Step 3: As shown in Fig. 9, the trained FIRE-DET model M predicts the frame and type of the flame object, and then obtains and preserves the flame location information; Visualize the preserved flame target position in the original moving image; After the ratio of the sum of the target area of each flame in the video frame and the area of the original video frame reaches a predetermined threshold, a flame alarm is issued.

A moving picture frame image IMG is obtained, and preprocessing is performed to obtain IMG0. The image IMG1 is obtained by normalizing the size of the moving picture frame image IMG0. This embodiment is normalized to a size of 416 × 416 pixels and is used as an input for the next stage model M.

The image of the normalized size is transferred into the model M, the coordinate position of the flame corresponding to the video frame image is stored in the array Loc, and the flame target position is marked in the video frame IMG1 after preprocessing and normalization based on the Loc to identify the flame By obtaining the effect diagram IMG2, the visualization of flame tracking is realized.

After the area occupied by the flame target in the video frame in the video stream reaches a set threshold, a flame alarm is sent. If the total area occupied by all flame targets among video frames is greater than 50%, a flame alarm is sent, otherwise the judgment is continued. Specifically, in the video frame image, one binary image FLAG is generated based on the flame position of Loc, and the flame position of the binary image FLAG is set to 1 and the other parts are set to 0; formula below,

에 의해 판단하며,

judged by

where Fire _s is the sum of all flame target areas, w is the width of the video frame and h is the height of the video frame. If the above formula is satisfied, a flame alarm is sent.

The present invention further discloses a lightweight FIRE-DET flame detection system, mainly comprising: an image pre-processing module used to read a moving picture frame image and pre-process and normalize the moving picture frame image; a flame detection module used to obtain location information of a flame target by detecting the video frame image after normalization using the trained FIRE-DET model; a flame zone visualization module used to mark the corresponding flame target in the original video image based on the flame target position preserved by the flame detection module and implement the visualization of flame tracking; and a flame alarm module: a flame alarm module used to continuously monitor the video and send a flame alarm to notify the user when the area occupied by the flame target among video frames reaches a predetermined threshold value. The system can implement the light-weight FIRE-DET flame detection method and belongs to the same invention concept, and specific details refer to the embodiment of the method, which will not be repeatedly described herein.

The lightweight FIRE-DET flame detection system disclosed by the embodiment of the present invention based on the equivalent inventive concept includes a memory, a processor, and a computer program that is stored in the memory and can be operated in the processor, and the computer program is loaded into the processor Implement the light-weight FIRE-DET flame detection method.

Claims (7)

In the light-weight FIRE-DET flame detection method,
(1) build a FIRE-DET model; the FIRE-DET model includes a feature extraction network, a feature fusion network, an image segmentation network and a predictive identification network; The feature extraction network is constructed by stacking multi-convolution combination structures, and reduces the number of convolutional channels to lower the parameter amount of the detection model; the feature fusion network is a BiFPN network; The image segmentation network performs deconvolution and convolution manipulation on the features after the fusion, then obtains an attention diagram, and fuses the features with the features obtained by the feature fusion network;
(2) forming a data set by performing pre-processing and normalization on a video frame image including a pre-obtained complex environment; In addition, training the FIRE-DET model to obtain a model M used to detect the flame;
(3) using the model M to predict the frame and type of the flame object, and then obtain and preserve the location information of the flame; Visualize the preserved flame target position in the original moving image; After the ratio of the sum of the target area of each flame in the video frame and the area of the original video frame reaches a predetermined threshold, sending a flame alarm; Lightweight FIRE-DET flame detection method comprising the. According to claim 1,
The operation process of the feature extraction network described in step (1) is,
Normalizes the input drawing to a size of 416×416 to generate IMG0; IMG0 after normalization is taken as an input of the multi-convolutional combination network and calculated to obtain F1; After performing a maximum pooling operation with a convolution kernel of 2×2 on F1, the obtained Pool1 is used as an input to the multi-convolutional combination network and calculated to obtain F2; After performing the maximum pooling operation with the convolutional kernel of 2×2 on F2, the obtained Pool2 is used as an input to the multi-convolutional combination network and calculated to obtain F3; After performing the maximum pooling operation with the convolutional kernel of 2×2 on F3, the obtained Pool3 is used as an input to the multi-convolutional combination network and calculated to obtain F4; After performing the maximum pooling operation with a convolutional kernel of 2×2 on F4, the obtained Pool4 is used as an input to the multi-convolutional combination network and calculated to obtain F5; After performing a maximum pooling operation with a convolutional kernel of 2×2 on F5, a convolution operation with a size of a convolutional kernel of 3×3 is performed to obtain F6; Light-weight FIRE-DET flame detection, characterized in that after performing the maximum pulling operation with a convolutional kernel of 2×2 on F6, a convolution operation with a size of a convolutional kernel of 3×3 is performed to obtain F7; Way. According to claim 1,
The operation process of the feature fusion network described in step (1) is,
The features obtained through the multi-convolutional combination structure of the feature extraction network also take F3, F4, F5, F6 and F7 as inputs of the BiFPN network, fuse the features, and then obtain outputs C1, C2, C3, C4 and C5;
Taking C1, C2, C3, C4 and C5 as inputs of the BiFPN network, fusing the features again to obtain outputs D1, D2, D3, D4 and D5; The process of the BiFPN network is,
1) In the network, there are 5 inputs denoted as Input1, Input2, Input3, Input4 and Input5, respectively;
2) After performing deconvolution operation of 2×2 on Input1, arrange and operate with Input2 to obtain A1; After performing a 2×2 deconvolution operation on A1, arrange and operate with Input3 to obtain A2; After performing a 2×2 deconvolution operation on A2, arrange and operate with Input2 to obtain A3;
3) After performing 2×2 deconvolution operation on A3, arrange with Input5 and operate to obtain B5; After performing the maximum pulling operation of 2×2 on B5, arrange and operate with Input4 and A3 to obtain B4; After performing the maximum pulling operation of 2×2 on B4, arrange and operate with Input3 and A2 to obtain B3; After performing the maximum pulling operation of 2×2 on B3, arrange and operate with Input2 and A1 to obtain B2; After performing the maximum pulling operation of 2×2 on B2, arrange and operate with Input1 to obtain B1;
4) A light-weight FIRE-DET flame detection method characterized by using B1, B2, B3, B4 and B5 in each of 1) to 3) as the output of the feature fusion network. According to claim 1,
The implementation process of step (2) is,
(21) Extract the video frame set from the video by applying the method of adopting a key frame to the published flame video, and mark the location of the flame using the image marking tool on the video frame set and build a tag data set and;
(22) The binary image corresponding to each moving image in the training data set follows the tag data set, and the position of the flame indicated by the original drawing corresponding to the binary image is set to 1. However, other parts are set to 0 to form the tag data set of FIRE-DET's reason image, and finally, the data set of FIRE-DET is built into the video frame set, the flame target tag data set and the tag data set of the logical image. and;
(23) Increase the target pixel value based on the normal distribution of the moving picture frame image among the moving picture frame sets, and increase the data of the data set by randomly generating horizontal mirroring for the moving picture frame image. Lightweight class FIRE-DET flame detection method. According to claim 1,
The loss function of the image segmentation network described in step (3) is

is,
where x is the attention diagram output by the image segmentation network, z is the rational image representing the flame target; A light-weight FIRE-DET flame detection method, characterized in that each of the three attention drawings output by the image segmentation network performs a loss function and a feedback calculation using a rational image to represent the flame. an image preprocessing module used to read a moving picture frame image and perform preprocessing and normalization on the moving picture frame image;
a flame detection module used to obtain location information of a flame target by detecting a video frame image after normalization using the trained FIRE-DET model;
a flame area visualization module used to realize the visualization of flame tracking by marking the corresponding flame target in the original video image based on the flame target position preserved by the flame detection module;
Any one of claims 1 to 4 comprising a; flame alarm module used to continuously monitor the video and send a flame alarm to notify the user when the flame target exceeds the predetermined area in the video frame A lightweight FIRE-DET flame detection system according to the method. A light-weight FIRE-DET flame detection method according to any one of claims 1 to 5, including a memory, a processor, and a computer program stored in the memory and capable of running in the processor, wherein the computer program is loaded to the processor A lightweight FIRE-DET flame detection system characterized in that it is implemented.

Images