Forest fire image segmentation method and device based on overhead transmission line channel monitoring image

Abstract

The invention provides a mountain fire image segmentation method and device based on an overhead transmission line channel monitoring image. The method comprises the following steps: extracting a flame candidate region from a visible light image of a mountain fire; enhancing the image edge of the flame candidate area by using a spatial position distance and a color channel distance; and carrying out category judgment on edge points of the image edge of the enhanced flame candidate region, and finding out the region where the mountain fire image is located. The method achieves the image marking of the forest fire of the power transmission line channel, improves the accuracy of flame image segmentation prediction under a complex background, and has the advantage of wide applicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission line operation and maintenance disaster prevention early warning, in particular to a mountain fire image segmentation method and device based on overhead power transmission line channel monitoring images. BACKGROUND

[0002] Overhead power transmission lines are an important part of China's power grid, playing an important role in transmitting electric energy and connecting power plants and substations. Due to the vast territory of China, overhead power transmission lines inevitably pass through mountainous areas with dense trees. If a mountain fire hazard occurs near the overhead power transmission line channel, the mountain fire and smoke may cause the overhead power transmission line to have insufficient safety distance and trip, or the insulator performance to fail, resulting in serious accidents and endangering the safe and stable operation of the power system.

[0003] Traditional monitoring methods include different methods based on manual inspection, digital image technology, and deep learning. The mountain fire monitoring method based on manual inspection is to have staff regularly patrol the overhead power transmission line channel to check whether a mountain fire has occurred. This method is inefficient, especially when the mountain fire spreads quickly and widely, making it impossible to identify the mountain fire hazard in time. The mountain fire monitoring method based on digital image technology is to extract the features of the mountain fire from the video images by using cameras or video monitoring the overhead power transmission line channel. Due to the consideration of a single color space model, the method is greatly affected by the image background environment and has poor accuracy, which cannot meet the requirements of power transmission line operation and maintenance disaster prevention early warning. The mountain fire monitoring method based on deep learning uses a deep learning network to infer the monitoring images and segment the mountain fire area for judgment. Since the training of the deep learning network relies on a large number of mountain fire images, and the actual number of mountain fire images of the overhead power transmission line channel is small, the generalization ability of the trained network is insufficient, which may lead to missed reports and false reports. SUMMARY

[0004] The purpose of the present application is to provide a mountain fire image segmentation method and device based on overhead power transmission line channel monitoring images, which improves the accuracy of mountain fire image recognition and has the advantage of wide applicability.

[0005] To achieve the above purpose, the present application provides a mountain fire image segmentation method based on overhead power transmission line channel monitoring, comprising the following steps:

[0006] extracting a flame candidate region from a visible light image of the mountain fire;

[0007] enhancing the image edges of the flame candidate region using spatial position distance and color channel distance;

[0008] performing category determination on the edge points of the enhanced image edges of the flame candidate region to find the region where the mountain fire image is located.

[0009] Furthermore, extracting the candidate flame area from the visible light image of the wildfire includes:

[0010] (1) Convert the visible light image of the wildfire from RGB color space to Lab color space and separate it into the brightness channel (L) and two color channels (a and b);

[0011] (2) Gaussian filtering is performed on the brightness channel (L) and two color channels (a and b) in the obtained Lab color space to denoise them, and then normalized using linear transformation;

[0012] (3) Calculate the average value of the brightness channel (L) as the threshold adjustment factor ;

[0013] (4) Calculate the histogram of the color channel (b) and take the peak value in the histogram as and threshold adjustment factor Combined to get the threshold range [ p b × ( 1 . 1 + 0 . 5 f L 255 ) , 255 ] , preliminary segmentation of candidate regions in the image based on the threshold range;

[0014] (5) Use the closing operation of the morphological operator to eliminate the holes in the candidate area and select the flame candidate area.

[0015] Furthermore, the method of enhancing the image edge of the flame candidate region by using the spatial position distance and the color channel distance includes:

[0016] (1) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Spatial distance and weights ,in is the variance of spatial position distance;

[0017] (2) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Color channel (b) distance and weights ,in is the variance of the color channel distance;

[0018] (3) Pixels The pixel values ​​in the neighborhood of are weighted summed to obtain the enhancement value of the point I e ( x , y ) = I ( x , y ) + α [ I ( x , y ) − ∑ w s ( i , j ) × w b ( i , j ) × I ( x , y ) ∑ w s ( i , j ) × w b ( i , j ) ] .

[0019] Furthermore, the edge points of the image edge of the enhanced flame candidate area are subjected to category determination, which is characterized by:

[0020] (1) Calculate the gradient amplitude of the image after enhancing the edge , the gradient direction , and the maximum inter-class variance threshold of the gradient ;

[0021] (2) If the gradient of a certain pixel point in the image after enhancing the edge is greater than the gradient of the adjacent pixel point in the gradient direction and the gradient of the pixel point , and the pixel point and have and only one point is a possible smoke point, then the pixel point is a flame edge, otherwise it is not a flame edge;

[0022] (3) From the flame edge, search whether there is a pixel point with a gradient greater than the threshold in its 8-neighborhood. If there is, then the point is connected with the flame edge, and the same search is performed. If there is not, the final flame edge is obtained.

[0023] Further, the judgment method of the pixel point being a possible smoke point is:

[0024] (1) Use the color channel (R), color channel (G), and color channel (B) values of the point in the RGB color space to determine;

[0025] (2) If the variance of the RGB three values of the point is less than 80, and the mean value range is between [100, 200], it is a possible smoke point; otherwise, it is not.

[0026] An image segmentation device for mountain fire based on overhead transmission line channel monitoring, comprising:

[0027] a flame candidate region extraction module for extracting a flame candidate region from a visible light image of a mountain fire;

[0028] an image edge enhancement module for enhancing the image edge of the flame candidate region using spatial position distance and color channel distance;

[0029] a mountain fire image recognition module for classifying the edge points of the enhanced image edge of the flame candidate region and finding the region where the mountain fire image is located.

[0030] Further, the flame candidate region extraction module extracts a flame candidate region from a visible light image of a mountain fire, comprising:

[0031] (1) Convert the visible light image of the wildfire from RGB color space to Lab color space and separate it into the brightness channel (L) and two color channels (a and b);

[0032] (2) Gaussian filtering is performed on the brightness channel (L) and two color channels (a and b) in the obtained Lab color space to denoise them, and then normalized using linear transformation;

[0033] (3) Calculate the average value of the brightness channel (L) as the threshold adjustment factor ;

[0034] (4) Calculate the histogram of the color channel (b) and take the peak value in the histogram as and threshold adjustment factor Combined to get the threshold range [ p b × ( 1 . 1 + 0 . 5 f L 255 ) , 255 ] , preliminary segmentation of candidate regions in the image based on the threshold range;

[0035] (5) Use the closing operation of the morphological operator to eliminate the holes in the candidate area and select the flame candidate area.

[0036] Furthermore, the image edge enhancement module uses the spatial position distance and the color channel distance to enhance the image edge of the flame candidate area, including:

[0037] (1) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Spatial distance and weights ,in is the variance of spatial position distance;

[0038] (2) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Color channel (b) distance and weights ,in is the variance of the color channel distance;

[0039] (3) Pixels The pixel values ​​in the neighborhood of are weighted summed to obtain the enhancement value of the point I e ( x , y ) = I ( x , y ) + α [ I ( x , y ) − ∑ w s ( i , j ) × w b ( i , j ) × I ( x , y ) ∑ w s ( i , j ) × w b ( i , j ) ] .

[0040] Furthermore, the wildfire image recognition module determines the category of edge points of the image edge of the enhanced flame candidate area to find the area where the wildfire image is located, including:

[0041] (1) Calculate the image gradient amplitude after edge enhancement , gradient direction and the maximum inter-class variance threshold of the gradient ;

[0042] (2) If a pixel in the image after edge enhancement Gradient Greater than its gradient direction The adjacent pixels on Gradient and pixels Gradient , and the pixel and If there is only one possible smoke point, then the pixel It is the edge of the flame, otherwise it is not the edge of the flame;

[0043] (3) Starting from the edge of the flame, search its 8 neighborhoods to see if there is a gradient greater than the threshold If it exists, the point is connected to the flame edge and the same search is performed. If it does not exist, the final flame edge is marked.

[0044] Furthermore, the method for determining whether the pixel point is a possible smoke point is:

[0045] (1) Use the color channel (R), color channel (G), and color channel (B) values ​​of the point in the RGB color space to make a judgment;

[0046] (2) If the variance of the RGB values ​​of the point is less than 80 and the mean range is between [100, 200], it is a possible smoke point; otherwise, it is not.

[0047] The advantages of the present invention are:

[0048] 1. This method uses the color channel (b) in the Lab color space to better represent the color and chromaticity information of flames. The core colors of the flame (yellow, orange) correspond to the high positive values ​​of the color channel (b), while the background, such as the sky and soil, corresponds to the low positive or negative values ​​of the color channel (b). At the same time, the information of the luminance channel (L) is used to adjust the selection range of the flame candidate region, enhancing the algorithm's ability to adapt to different lighting conditions, thereby improving the accuracy of the flame candidate region selection.

[0049] 2、The edge information of the flame candidate region is enhanced using spatial position information and color channel information. Only when the pixel point is close to the pixel point and the color of the pixel point is greatly different from the flame feature, the pixel point is enhanced by weight. Thus, the possible flame edge information is highlighted, and the accuracy of flame edge discrimination is improved.

[0050] 3、The present application detects the possible flame edge in the gradient direction according to the feature that the flame and the smoke exist in a boundary, finds out when one side is a flame region and the other side is a smoke region, and then performs edge connection and thinning operation, so that the real boundary of the flame can be more accurately extracted. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a flowchart of the mountain fire image recognition method of the overhead power transmission line channel of the present application.

[0052] Figure 2 It is a detection case of the mountain fire image recognition method of the overhead power transmission line channel of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the present application will be described clearly and completely in combination with the drawings in the present application.

[0054] As shown in the drawings, Figure 1 The present application provides a mountain fire image recognition method of an overhead power transmission line channel, which comprises the following steps:

[0055] Step P101: converting the obtained visible light image from RGB color space to Lab color space, and extracting a flame candidate region;

[0056] Specifically, 1) converting the visible light image of the mountain fire from RGB color space to Lab color space, and separating to obtain a brightness channel (L) and two color channels (a and b), wherein the RGB color space is first converted to XYZ color space,

[0057] [ X Y Z ] = [ 0 . 4123 0 . 3577 0 . 1802 0 . 2126 0 . 7153 0 . 0721 0 . 0193 0 . 1192 0 . 9493 ] [ R G B ]

[0058] and then converting the XYZ color space to Lab color space,

[0059] { L = 116 f ( Y ) a = 500 [ f ( X 0 . 950456 ) − f ( Y ) ] b = 200 [ f ( Y ) − f ( Z 1 . 088754 ) ] , wherein ;

[0060] 2) Gaussian filtering is performed on the obtained brightness channel (L) and two color channels (a and b) in the Lab color space, and then linear change is used for normalization to reduce the influence of noise and uneven illumination, wherein the Gaussian filtering function is

[0061] ;

[0062] 3) Calculate the average value of the brightness channel (L) as the threshold adjustment factor,

[0063] ;

[0064] 4) Calculate the histogram of the color channel (b) and take the peak value in the histogram Combined with the threshold adjustment factor, the threshold range is obtained [ p b × ( 1 . 1 + 0 . 5 f L 255 ) , 255 ] ;

[0065] 5) Use the closing operation of morphological operators to eliminate holes in the region and select candidate wildfire areas.

[0066] Step P102: using the spatial position distance and the color channel distance to enhance the image edge of the flame candidate region.

[0067] Specifically, the flame candidate region extracted in step (1) is enhanced to highlight the edge of the image. The specific steps are as follows:

[0068] 1) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Spatial distance and weights ,in is the variance of spatial position distance;

[0069] 2) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Color channel (b) distance and weights ,in is the variance of spatial position distance;

[0070] 3) Pixels The pixel values ​​in the neighborhood of are weighted summed to obtain the enhancement value of the point I e ( x , y ) = I ( x , y ) + α [ I ( x , y ) − ∑ w s ( i , j ) × w b ( i , j ) × I ( x , y ) ∑ w s ( i , j ) × w b ( i , j ) ] .

[0071] Step P103: performing category determination on the edge points of the image edge of the enhanced flame candidate area to find the area where the wildfire image is located.

[0072] The specific steps are:

[0073] 1) Calculate the image gradient amplitude after edge enhancement , gradient direction and the maximum inter-class variance threshold of the gradient ;

[0074] 2) If a pixel Gradient Greater than the adjacent pixel in its gradient direction Gradient and pixels Gradient , and the pixel and If there is only one possible smoke point, then the pixel It is the edge of the flame, otherwise it is not the edge of the flame;

[0075] When determining a possible smoke point, the color channel (R), color channel (G), and color channel (B) values ​​of the point in the RGB color space are used. If the variance of the RGB values ​​of the point is less than 80 and the mean range is between [100, 200], it is a possible smoke point; otherwise, it is not.

[0076] 3) Starting from the edge of the flame, search its 8 neighborhoods for a gradient greater than the threshold If it exists, the point is connected to the flame edge and the same search is performed again. If it does not exist, the final flame edge is marked.

[0077] Figure 2 For the case of detecting wildfire images according to the above method, complete and accurate extraction of flame edges was achieved.

[0078] An embodiment of the present invention further provides a wildfire image segmentation device based on overhead power line channel monitoring, comprising:

[0079] The flame candidate region extraction module is used to extract flame candidate regions from visible light images of wildfires;

[0080] An image edge enhancement module, configured to enhance the image edge of the flame candidate region using spatial position distance and color channel distance;

[0081] The wildfire image recognition module is used to determine the categories of edge points on the edge of the image of the enhanced flame candidate area and find the area where the wildfire image is located.

[0082] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by technicians in this technical field within the technical scope disclosed in the present invention should be covered by the scope of protection of the present invention.

Claims

1. A wildfire image segmentation method based on overhead transmission line channel monitoring images, characterized in that: The steps include: Extracting flame candidate regions from visible light images of wildfires; Using spatial position distance and color channel distance to enhance the image edge of the flame candidate area; The edge points of the enhanced flame candidate area are classified to find the area where the wildfire image is located.

2. The method for segmenting wildfire images based on overhead transmission line channel monitoring images according to claim 1, characterized in that: Extracting a candidate flame region from a visible light image of a wildfire includes: (1) Convert the visible light image of the wildfire from RGB color space to Lab color space and separate it into the brightness channel (L) and two color channels (a and b); (2) Gaussian filtering is performed on the brightness channel (L) and two color channels (a and b) in the obtained Lab color space to denoise them, and then normalized using linear transformation; (3) Calculate the average value of the brightness channel (L) as the threshold adjustment factor ; (4) Calculate the histogram of the color channel (b) and take the peak value in the histogram as and threshold adjustment factor Combined to get the threshold range , preliminary segmentation of candidate regions in the image based on the threshold range; (5) Use the closing operation of the morphological operator to eliminate the holes in the candidate area and select the flame candidate area.

3. The method for segmenting wildfire images based on overhead transmission line channel monitoring images according to claim 1, characterized in that: The method of enhancing the image edge of the flame candidate region by using the spatial position distance and the color channel distance includes: (1) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Spatial distance and weights ,in is the variance of spatial position distance; (2) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Color channel (b) distance and weights ,in is the variance of the color channel distance; (3) Pixels The pixel values ​​in the neighborhood of are weighted summed to obtain the enhancement value of the point .

4. The method for segmenting wildfire images based on overhead transmission line channel monitoring images according to claim 1, characterized in that: The edge points of the enhanced flame candidate area are classified to find the area where the wildfire image is located, including: (1) Calculate the image gradient amplitude after edge enhancement , gradient direction and the maximum inter-class variance threshold of the gradient ; (2) If a pixel in the image after edge enhancement Gradient Greater than its gradient direction The adjacent pixels on Gradient and pixels Gradient , and the pixel and If there is only one possible smoke point, then the pixel It is the edge of the flame, otherwise it is not the edge of the flame; (3) Starting from the edge of the flame, search its 8 neighborhoods to see if there is a gradient greater than the threshold If it exists, the point is connected to the flame edge and the same search is performed. If it does not exist, the final flame edge is marked.

5. The method for segmenting wildfire images based on overhead transmission line channel monitoring images according to claim 4, characterized in that: The method for determining whether the pixel point is a possible smoke point is: (1) Use the color channel (R), color channel (G), and color channel (B) values ​​of the point in the RGB color space to make a judgment; (2) If the variance of the RGB values ​​of the point is less than 80 and the mean range is between [100, 200], it is a possible smoke point; otherwise, it is not.

6. A wildfire image segmentation device based on overhead transmission line channel monitoring, characterized in that: include: The flame candidate region extraction module is used to extract flame candidate regions from visible light images of wildfires; An image edge enhancement module, configured to enhance the image edge of the flame candidate region using spatial position distance and color channel distance; The wildfire image recognition module is used to determine the categories of edge points on the edge of the image of the enhanced flame candidate area and find the area where the wildfire image is located.

7. The wildfire image segmentation device based on overhead power line channel monitoring images according to claim 6, characterized in that: The flame candidate region extraction module extracts flame candidate regions from the visible light image of the wildfire, including: (1) Convert the visible light image of the wildfire from RGB color space to Lab color space and separate it into the brightness channel (L) and two color channels (a and b); (2) Gaussian filtering is performed on the brightness channel (L) and two color channels (a and b) in the obtained Lab color space to denoise them, and then normalized using linear transformation; (3) Calculate the average value of the brightness channel (L) as the threshold adjustment factor ; (4) Calculate the histogram of the color channel (b) and take the peak value in the histogram as and threshold adjustment factor Combined to get the threshold range , preliminary segmentation of candidate regions in the image based on the threshold range; (5) Use the closing operation of the morphological operator to eliminate the holes in the candidate area and select the flame candidate area.

8. The wildfire image segmentation device based on overhead power line channel monitoring images according to claim 6, characterized in that: The image edge enhancement module uses spatial position distance and color channel distance to enhance the image edge of the flame candidate area, including: (1) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Spatial distance and weights ,in is the variance of spatial position distance; (2) Calculate the pixel points in the flame candidate area With other pixels in the neighborhood Color channel (b) distance and weights ,in is the variance of the color channel distance; (3) Pixels The pixel values ​​in the neighborhood of are weighted summed to obtain the enhancement value of the point 。 9. The wildfire image segmentation device based on overhead power line channel monitoring images according to claim 6, characterized in that: The wildfire image recognition module determines the category of edge points of the image edge of the enhanced flame candidate area to find the area where the wildfire image is located, including: (1) Calculate the image gradient amplitude after edge enhancement , gradient direction and the maximum inter-class variance threshold of the gradient ; (2) If a pixel in the image after edge enhancement Gradient Greater than its gradient direction The adjacent pixels on Gradient and pixels Gradient , and the pixel and If there is only one possible smoke point, then the pixel It is the edge of the flame, otherwise it is not the edge of the flame; (3) Starting from the edge of the flame, search its 8 neighborhoods to see if there is a gradient greater than the threshold If it exists, the point is connected to the flame edge and the same search is performed. If it does not exist, the final flame edge is marked.

10. The wildfire image segmentation device based on overhead power line channel monitoring images according to claim 9, characterized in that: The method for determining whether the pixel point is a possible smoke point is: (1) Use the color channel (R), color channel (G), and color channel (B) values ​​of the point in the RGB color space to make a judgment; (2) If the variance of the RGB values ​​of the point is less than 80 and the mean range is between [100, 200], it is a possible smoke point; otherwise, it is not.

Images