CN113920417A

CN113920417A - Open fire detection method based on Gauss YOLOv3

Info

Publication number: CN113920417A
Application number: CN202111188172.2A
Authority: CN
Inventors: 宋长青; 汪创; 曾耀强; 吴彦伟; 姚丙君
Original assignee: Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd
Current assignee: Guangzhou Power Supply Bureau of Guangdong Power Grid Co Ltd
Priority date: 2021-10-12
Filing date: 2021-10-12
Publication date: 2022-01-11

Abstract

The invention discloses an open fire detection method based on Gauss YOLO v 3. The method comprises the following steps that a camera collects a field image of the transformer substation containing open fire; performing marking processing by using the bounding box to establish an image data set; detecting and positioning open fire in the transformer substation field image through the modified Gauss YOLO v3 model after the image is zoomed to obtain a detection model, and testing the effect of the model by using a verification set; and inputting the transformer substation field image acquired in real time, and processing the transformer substation field image through a detection model to obtain the position coordinates of the open fire. The method solves the problem of uncertainty estimation of the open fire detection frame, and the identification result of the transformer substation site shows that the identification method has higher accuracy and can be applied to an identification system for open fire detection in the transformer substation construction site.

Description

Open fire detection method based on Gauss YOLOv3

Technical Field

The invention relates to a method for detecting open fire of a transformer substation, in particular to an open fire detection method based on Gauss YOLO v 3.

Background

With the large-scale development of electric power systems in China, the number of substations is increased rapidly, and the number of 35 kilovolt substations reaches more than 25000. However, the number of people in the transformer substation in China is less than 3, and the serious shortage is up to more than 78%. Therefore, fire safety accidents caused by high temperature of equipment occur, and the problems of increased fire safety risk of the transformer substation, insufficient human resources, insufficient refinement and the like are urgently needed to be solved, so that the research and development of an automatic naked fire detection and identification method are urgently needed to reduce the safety risk.

Disclosure of Invention

In order to solve the problems in the background art, the invention provides an open fire detection method based on Gauss YOLO v3, which can efficiently identify the open fire of a transformer substation and has good stability.

The technical scheme of the invention comprises the following steps:

1) shooting and collecting a transformer substation field image containing open fire through a camera (as shown in figure 1);

2) traversing all the site images of the transformer substation, marking each image by using a bounding box aiming at open fire in the image to obtain a corresponding label file, and enabling the label file and the original site image of the transformer substation to jointly form an image data set;

3) carrying out image scaling on a transformer substation field image in an image data set, adjusting the image size to be 416 x 416, carrying out same-scale scaling on a corresponding label file, detecting and positioning naked fire in the transformer substation field image through a modified YOLO v3 model to obtain a detection model, and finally testing the model effect by using a verification set;

4) and (3) zooming the live images of the transformer substation acquired in real time into 416 x 416 according to the same image zooming method as the step 3), and obtaining the position coordinates of the open fire through the processing of the detection model as the input of the detection model.

The field image of the transformer substation in the step 1) refers to an image acquired by a camera in the transformer substation.

The bounding boxes in the step 2) are rectangular boxes, are expressed into 1 × 4 row vectors, and include horizontal and vertical coordinates corresponding to the upper left corner and the lower right corner of the rectangular boxes, the horizontal and vertical coordinates of the bounding boxes are recorded in a label file, each bounding box corresponds to one label file, and each substation field image only contains one bounding box at most.

In the step 3), the regression of the coordinate frame in the modified YOLO v3 model is implemented by gaussian modeling, so that a gaussian YOLO v3 model is formed, and for each coordinate of the regression frame, the regression value includes a mean value and a variance, the mean value reflects the mean value of the frame coordinates, and the variance reflects the uncertainty of the mean value.

And finally, normalizing the variance of the output of the detection model to be between 0 and 1 through a sigmoid function.

In the detection model, in the non-maximum suppression operation, the score of the detection frame is the product of the category score and the average variance of the four-side frame of the detection frame.

According to the method, YOLO v3 is modified, the traditional Dirac modeling is replaced by Gaussian modeling in the coordinate regression process, so that the uncertainty of the detection frame is obtained, and the final frame is screened by combining the uncertainty of the position of the detection frame and the category confidence coefficient. For open fire detection, firstly, the size of a picture to be detected is adjusted to 416 x 416, the picture is input into Gauss YOLO v3, and an open fire detection frame is obtained through output.

The invention has the beneficial effects that:

the method solves the problem of substation open fire detection, and the identification result based on the substation scene shows that the accuracy of the identification method reaches 89.4%, is 8.63% higher than that of the original yolov3, can be applied to an identification system of substation open fire detection, and has universality for identification of other objects.

Drawings

Fig. 1 is an example data sample picture.

FIG. 2 is an example of an embodiment image dataset.

FIG. 3 is a sample image data set of the scaled embodiment.

FIG. 4 is a diagram showing the results of the detection model according to the embodiment of the present invention.

Detailed Description

The invention is described in further detail below with reference to the figures and the embodiments.

The examples of the invention are as follows:

firstly, a camera is used for shooting and collecting a field image of a substation containing an open fire, as shown in fig. 1, 5311 images are collected in the embodiment, the size of an original image is 1920 × 1080 pixels, 4246 images are used for training, and 1065 images are used as test images;

traversing all substation field images containing open fires, marking each image by using a bounding box aiming at the open fire in the image to obtain a corresponding label file, and forming an image data set by the label file and the original substation field image together as shown in fig. 2;

performing image scaling on a site image of a transformer substation in an image data set, adjusting the size of the image to 416 × 416, and scaling a corresponding label file in the same scale as that of the site image, as shown in fig. 3;

the modified YOLO v3 model was trained with an image dataset to obtain a detection model.

And (5) outputting the test set picture to a model, and detecting and positioning the open fire in the test set picture to obtain a detection result, as shown in fig. 4.

The detection model trained by the invention is adopted to detect the pictures of the field test set of the transformer substation, and the obtained results are shown in table 1. As can be seen from table 1, the average accuracy of open flame identification was 89.4%, which is an 8.63% improvement over the original yolov 3.

TABLE 1

YOLOv3	82.73％
		Gauss YOLOv3	91.36％

Therefore, the method can realize automatic identification of the open fire of the transformer substation, has high accuracy, and has the advantages of good stability, strong anti-interference capability, high universality and the like.

The foregoing detailed description is intended to illustrate and not limit the invention, which is intended to be within the spirit and scope of the appended claims, and any changes and modifications that fall within the true spirit and scope of the invention are intended to be covered by the following claims.

Claims

1. An open fire detection method based on Gauss YOLO v3 is characterized by comprising the following steps:

1) shooting and collecting a transformer substation field image containing open fire through a camera;

3) carrying out image scaling on a transformer substation field image in an image data set, carrying out same-scale scaling on a corresponding label file, detecting and positioning naked fire in the transformer substation field image through a modified YOLO v3 model to obtain a detection model, and finally testing the effect of the model by using a verification set;

4) and (3) zooming the transformer substation field image acquired in real time according to the same image zooming method as the step 3), and obtaining the position coordinates of the open fire through the processing of the detection model as the input of the detection model.

2. The open fire detection method based on the Gaussian YOLO v3 as claimed in claim 1, wherein: the field image of the transformer substation in the step 1) refers to an image acquired by a camera in the transformer substation.

3. The open fire detection method based on the Gaussian YOLO v3 as claimed in claim 1, wherein: the bounding boxes in the step 2) are rectangular boxes, are expressed into 1 × 4 row vectors, and include horizontal and vertical coordinates corresponding to the upper left corner and the lower right corner of the rectangular boxes, the horizontal and vertical coordinates of the bounding boxes are recorded in a label file, each bounding box corresponds to one label file, and each substation field image only contains one bounding box at most.

4. The open fire detection method based on the Gaussian YOLO v3 as claimed in claim 1, wherein: in the step 3), the coordinate frame regression in the modified YOLO v3 model is gaussian modeled, so that a gaussian YOLO v3 model is formed, and for each coordinate of the regression frame, the regression value includes a mean and a variance.

5. The open fire detection method based on the Gaussian YOLO v3 as claimed in claim 1, wherein: and finally, normalizing the variance of the output of the detection model to be between 0 and 1 through a sigmoid function.

6. The open fire detection method based on the Gaussian YOLO v3 as claimed in claim 1, wherein: in the detection model, in the non-maximum suppression operation, the score of the detection frame is the product of the category score and the average variance of the four-side frame of the detection frame.