TW200834476A - A fire image detection method - Google Patents

Abstract

A fire image detection method is used to compare a fire image sample with a series of real-time caught image frames and decide whether a fire breaks out in the image-caught scene or not. The method includes steps of converting the fire image sample into a specific color space, calculating a central point of the fire image sample, deciding a threshold value by the methods of cluster analysis and Mahalanobis distance, calculating a pixel distance between the central point and each pixel of one of real-time caught image frames by the method of Mahalanobis distance, deciding that a fire breaks out in the scene if the pixel distance is smaller than the threshold value.

Description

200834476 IX. Description of the invention: [Technical field of invention] A fire image_method, in particular, a similarity analysis of an image pixel with a fire and a k-paper to know how close the image pixel is to the fire sample. Whether there is a fire in the image. [Prior Art] In the technique of judging whether there is a fire after using the image to capture the surface, we can see that the United States still has the 3722 scale "Flredetecti〇nsystem", which is first through the ones, the other, the work/storage, and the outside line. The sensory person senses the illuminated area. If the detected energy exceeds the set threshold (Thresh〇ld), the image is captured by the camera, and the image is analyzed for the brillllt areas. Edge detection and image s^traction, etc., to determine whether it is a real fire. Under this architecture, the detection of fire requires the use of an additional sensor to assist in the more accurate determination of the range of vibrations to be sensed. In addition, in Wen-Bing Horng, Jian-Wen Peng, Chih-Yuan Chen 5 200834476, 2005 American Institute of Electrical and Electronics Engineers Yang EE, March 19-22, 2005) paper "A new image-based real- Time flame detection method using color anaiysis", the detection process is divided into two parts: static and dynamic. In static aspect, the fire samples are collected in advance, and the samples are red, green, blue, and RGB domain. After converting to HSI color space (HIS domain, Hue, Saturation, Luminance Intensity), the range of HSI values of sample fire and smoke is counted, and these statistics are used to identify the process of fire and smoke. On the dynamic side, the image difference method is used to remove background objects similar to fire and smoke to determine whether there is fire and smoke. Secondly, it was published in 2004·H·Chen, Ρ·H· Wu and Y·C. Chiou at the 2004 International Conference on Image Processing (ICIP, October 24, 27, 2004). An Early Fire-Detection Method Based on Image Processing, which divides the processing into two parts, static and dynamic, and the static part, which is classified in the RGB color space (such as r $ g > B), and then Carrying out the fire on the hue of the 6 200834476 preliminary classification, in terms of dynamics, using the % image difference to remove the background similar to fire - the flat paper, in the static energy of the ancient face, are using a large number of collected fire The source is 'and the Ted or the range on the fire (such as the fiber gamma) and the value of the Newton multi-ship boundary, "turn the best wire, and then go to the area where the fire is initially classified. In this way, the fire and tree sets are more domain-based, and the operation is complicated. SUMMARY OF THE INVENTION The present invention is based on the use of a fire image of the present invention, and the method of the present invention is based on the method of taking the ^^^^~ image surface and the fire image sample into the pixels in the image field. Whether there is a fire or a classification of most fire image samples, you can lightly (4) break the shadow ^ pixels without using an additional sensor for analysis to solve the above problem. According to the μ material method of the present invention, the fairy-mf 彡 pattern is pulled at least 1 to take an image (cafe geF 丽 e) for comparison, to determine that the operation has a face like a fire pixel, the fire image detection The test method package 200834476 includes: converting each sample pixel of the fire image sample into a color space, the color model having a first coordinate and a second coordinate, each sample pixel corresponding to the first coordinate and The second coordinate has a first coordinate value and a second coordinate value; calculating a center point of the first coordinate value and the second coordinate value of the sample pixels; and calculating each of the sample pixels of the fire image sample Obtaining a plurality of sample distance values from the Mahalanobis Distance of the center point; defining a threshold value according to the sample distance values; taking one of the plurality of image pixels of the captured image plane as the determination pixel; Calculating a Mahjong distance between the determination pixel and the center point to obtain a pixel distance; and comparing the pixel distance to the sample distance, and when the pixel distance is small 8 200834476 at the threshold value It is judged that the image plane of the x image has the fire pixel. By the above method, g can use a fire image sample and the captured image to make a judgment to determine whether or not there is a fire pixel to achieve the foregoing purpose. In order to make the objective of the present invention, the implementation method and the function thereof have a further understanding, the following is the following: [Embodiment] The first embodiment of the present invention is the fire image detecting method of the present invention. The flow chart mainly compares the fire image sample 3〇 (please refer to “Fig. 2”) with at least the captured image frame (4〇, image frame) to determine the captured image surface. Whether there is a fire pixel; the fire image detection method includes: Step S12: converting each sample pixel of the fire image sample 3 into a color space (Color Space), the color model having a first coordinate and a first a second block, each of the sample pixels having a first coordinate value and a second coordinate value corresponding to the first coordinate and the second coordinate; Step S14: calculating the first coordinate value and the second of the sample pixels Block 9 200834476 one of the center points of the calibration; step S16: calculating a Mahalanobis Distance of each of the sample pixels of the fire image sample and the center distance of the medium sample to obtain a plurality of values; Step S18 · According to the Looking for samples Deviation definition_gate value · Step S20: taking one of the plurality of image pixels of the captured image frame 4 as a determination pixel; Step S22: calculating a Mahalanobis distance of the determination pixel from the center point to obtain a pixel distance And step S24. Comparing the pixel distance with the sample distance, and determining that the captured image plane has the fire pixel when the pixel distance is less than the threshold value. Wherein, the color model in step S1.2 can be, for example, RGB, YIQ, CMy, HSV, HSI #, and the invention uses the luminance color difference color model YIQ as an example, wherein the first private (7) is brightness ([leg - Occupation) coordinates, the second sample (I) is the chromatic phase difference component (lnphase) coordinates, the Q coordinate is the chromatic aberration quadrature component (Quadrature), and the YIQ system is the National Television System Committee painting i 10 200834476

The Television System Committee (NTSC) stipulates that the present invention is exemplified by the luminance coordinate and the color difference and the same component coordinate, and the color model and the selection of the two coordinates are considered in various color models. In the luminance chromatic aberration color model, the luminance coordinate and the chrominance difference component coordinate are used in the practice of the present invention, and the threshold value is set relatively easily, and a fairly positive judgment result can be obtained. The definitions of the color models listed above are briefly described below, in which the CMY (Cyan Magenta Yellow) color model is the most commonly used color model in the color image printing industry. In the color cube, the three colors are red and green. The blue complementary color, the CMY color model is the three color to represent the various colors of the color model. The HSV (Hue, Saturation, Value) color model mainly corresponds to the color matching model of the family, which is consistent with the visual perception of the human eye. It is a model suitable for human eye recognition, in which Η defines the wavelength of color production, called color tone, S Defining the degree of lightness, called saturation, V defines the degree of lightness and darkness of the color, called brightness, usually expressed as a percentage. Secondly, in the method of converting sample pixels into a color difference color model, it is more common to be red and green 11 200834476

The RGB domain is converted to a YIQ domain, so the formula for this conversion is listed below:

Y 0299 0.587 ^ 0.114 RI — 0.595716 -0.274453 - 0.321263 GQ 0.211456 -0.522591 0.311135 B From the above formula, there is a high correlation between the YIQ color model and the RGB color model. The difference is that the YIQ coordinates represent brightness, The chromatic aberration is the same component and the color difference is orthogonal component, and RGB is the red, green and blue coordinates, respectively. After the sample pixels are converted, each sample pixel will have a first coordinate value and a second coordinate value corresponding to the Y, I coordinate, that is, the luminance value and the color difference same direction component value, of course, each sample pixel will also have Q. The color difference quadrature component value of the coordinate, but since this coordinate is not taken as an example in the present embodiment, it will not be described again. Secondly, the first coordinate value of the sample pixels and the center point of the second coordinate value are calculated in step S14, and the arithmetic value of the first coordinate value and the second coordinate value of all the sample pixels are respectively calculated. The number means that all the first coordinates 12 200834476 are added and the value of the pixel is added, and all the second coordinate values are added and divided by the number of sample pixels. Thus, the coordinate value of the center point is obtained. The Markov distance of the step sis is the distance method in the cluster analysis in the statistical method, and the calculation formula is: where 'Yes·(Wx3,··.. X, )τ, # is the coordinate value of the center pointΣ Then, the covariate matrix in the joint group, (4) is the Markov distance of 5, which is detailed after detailed calculation.

This step mainly calculates the horses of each sample pixel in the fire image sample and the center point to form a complex number. In this series, the basis for defining the threshold value in step S18 is different from the nucleus distribution. The town has the following definitions, including setting the threshold value to a value greater than %% of the lemon distance value, defining the threshold value as the median of the sample distance values, and setting the threshold value to the samples. The maximum value of the distance value, the value of the door sandalwood 200834476 is set as the average value of the scale sample distance value, and the shot (four) value is set to the average value of the sample distance values multiplied by the "safety factor" value set to the sample value. The average value plus the standard deviation of three times the value of the county (four)', the safety factor can be adjusted according to the actual situation, such as the mouth. However, the silk method of various values changes depending on the histogram distribution of the sample distance. For example, the method of distributing the normal value of m times is better. If the distribution mode is not known, the maximum value may be used, and the value may be larger than The threshold value of the ratio (such as 95%) and the ratio of the fire pattern in the A-shirt sample are used to define this threshold. Furthermore, the image of the step S20 is captured by the camera, and the image is composed of continuous faces (frai plus), and in order to achieve the effect, it is necessary to capture in one second. The image kneading surface is better than at least thirty frames in the human eye's vision; step S20 rhyme Na (four) squeaking, one of which is to sequentially image the female singer in the image 40 The pixel is extracted as the judgment pixel, and then the subsequent steps are performed. In this way, the judgment of the hoisting feed and the fire pixel is more than that of 14 200834476, which will make the system load heavier, and may also be due to the system performance and the number of images. As a result, the speed of the operation cannot keep up with the speed of the drawing in the image plane. The second method may adopt a plurality of image pixels in the image frame of the captured image that are different from the image of the previous image, which is the judgment pixel, meaning that the image image and the previous time point are currently captured. The captured U-pixels of the image are compared one by one to find a plurality of different image pixels, and then the different image pixels are set one by one as the judgment pixels for the subsequent steps. In step S22 and step (four) 24, the Mahalanobis distance of the judgment pixel and the center point is calculated and the pixel pitch is paid, and the pixel distance is compared with the sample distance.

The same-cluster, also _ broken pixels should be fire pixels. Off, the image taken from the pixel to the touch axis 昼 has a fire pixel, * the setting of the gatecast value is based on the fire shadow.

In the picture, the fire image 15 200834476 仏本u is the shell % mode 'this - fire image sample can be taken from any actual fire image, the sample pixel of the wire fire image sample has 69 * 104 pixels, that is, a total of 7176 After the pixel is converted from the RGB coordinate to the YIG coordinate, the γ-seat and the I-station display have a matrix of [69 χ 104], and the two matrices are arranged in the following manner. Since the data is quite large, This embodiment only sets a specific numerical value as an explanation of the embodiment: X,. 1...』7176 (1...^7176 where Υ is 79.1697, Ii is 81.8977 as the following example) After step S14, the center point is obtained as her (10), where 1 and iu ΐμ are actually calculated as 255.0835 and 40.8868 respectively. The covariate matrix in the joint group of Υ and I is lacking]

(7176 r /=1, -Meari^X丨-Mean)7 · Λ (7176) -1 V J and take X! as an example: 79.1697 81.8977 255.0835 40.8868

X 79.1697 81.8977 255.0835 40.8868 104χ 2.1291 -0.5984 -0.5984 0.1682 16 200834476 In addition to adding the covariate matrix in the joint group one by one Ζ 2 +·····.Σ71761 - get (= 7176-1) to get ' = 103x L2 °86 *°·7754" 1^0.7754 0.6762 _ where 'in order to make the whole operation process not divide by zero and generate "曰^ also, can add - unit matrix (2), then get the combined group of bran variable silver The array is: 1.2096 -0.7754]

10 X 0.7754 0.6772 Next, the foregoing steps S16, S18 are performed, with the implementation of the present invention, column 1 > The distance sample distance distribution is not normally assigned, and there are 95% two ‘

The value is less than 6, so the threshold value is set to 6. Thereafter, the captured image plane 40 is first taken (such as one of the plurality of image pixels in the "Fig. 3" as the judgment pixel, and it is assumed that the Y-station of the judgment pixel is the pixel and the c-point is Markov. The distance is calculated as follows

A

To" α°- "225.0835' 40.8868 ^ — r, r JX 103 X \ 1.2096 - 0.7754" :0.7754 0.6772 ^~/ X ^ l V -0 Meaning - "225.0835' ^ 40.8868 T\ 17 200834476 0 15.23, since its value is greater than the threshold value, it is judged that it is not a fire pixel, thus completing the judgment of an image pixel point. For the determination of the determination pixel, please refer to "Fig. 4", which is another image image 41 of the captured image, which can be seen to have a flame pixel 42 therein, and in the previous image surface 40 There is no such flame pixel. Therefore, when it is judged that the pixel is compared with the previous two images, 40, 41 are compared, and the changed image pixel is calculated by the Mahalanobis distance, and some operation time can be omitted, and the flame pixel is omitted. After calculating the Mahalanobis distance, the value is 3.25, which is less than the threshold value, that is, it is judged to be a fire pixel. Regarding the physical meaning of the foregoing method of the present invention, the cluster analysis (Classification) in the statistical method is used to directly calculate the distance from each fire pixel to the central value in the fire image sample by the Mahalanobis distance method, and After learning the Markov distance method to calculate the coordinates of the color space, the choice of brightness and chromatic aberration in the same direction component, or the luminance and chromatic aberration orthogonal components will have a better effect, therefore, at the threshold value After the definition, the judgment pixels in the newly captured image plane are calculated in the sample. The distance between the heart points of the 200834476 is less than the gates. The ^^^^p indicates that the judgment pixel is the same as the fire pixel in the fire image sample. Fresh $, hunting to judge the fire pixel. Please refer to "figure 5" which is a schematic diagram of a second embodiment of the fire image sample 31 of the present invention. The smoke image sample is taken as an example. Generally, in the event of a fire, smoke is usually accompanied, so the smoke image sample is also required. As an important basis for the judgment of fire pixels, the detailed method of the image and the foregoing are not repeated here. The rain image sample 3〇, 31 can also be a fire image sample and a smoke image sample. Therefore, the center point contains a fire center point and a smoke center point, and the sample distance includes the fire sample distance and - the smoke sample distance 'pixel distance system includes - fire pixel distance and a smoke pixel distance, the gate value includes - fire n 槛 value and - 褴 value, so that when the fire pixel distance is less than the fire threshold value and the smoke When the pixel distance is less than the threshold value of the smoke, it is judged that the captured image mask has a fire pixel, and thus the judgment is more rigorous, and it is determined that the fire pixel is generated by the fire and the smoke at the same time. The fire image sample 30 and the captured image surface 4, 41 are preferably implemented by the same image capturing system and hardware. The pixel is 19 200834476, to eliminate hardware and hardware. The difference between the two. For the application of the present invention, please refer to "Fig. 6", which combines the foregoing steps S12, S14, S16, S18 into step S50: sample analysis to set the threshold value, and step S20, S22 is step S52, Calculating the pixel distance of the pixel, and then comparing the pixel distance with the threshold value in step S54, if the pixel distance is less than the threshold value, recording the image surface as a fire surface, and starting counting (S55), After the image is finished, step S56 is performed to determine whether there are three consecutive images of the image surface, and if so, an alarm is issued (S58). If not, step S59 is performed, and if the image is displayed, When the fire is not in the face, the fire is counted back to zero. In this way, the alarm can be avoided. Of course, the design of the above 30 consecutive fire screen counts can be adjusted according to different needs. Twenty or more or less to meet actual needs. While the present invention has been described in its preferred embodiments, the present invention is not intended to limit the invention, and it is obvious to those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention. The scope of the invention is defined by the scope of the patent application of the Chinese Patent Application No. 2008. ~ [Simple description of the drawing] ~ Figure 1 is a flow chart of the fire image detecting method of the present invention; Fig. 2 is a schematic view of the first embodiment of the fire image sample of the present invention; Schematic diagram of a captured image; • Figure 4 is a schematic view of another captured image; Figure 5 is a schematic view of a second embodiment of the fire image sample of the present invention; and The figure is a flow chart of an application example of the present invention. [Main component symbol description] 30, 31 Fire image sample 40, 41 Image 昼. Face 42 Flame pixel Step S12 Convert each sample pixel of the fire image sample into a color space (Step S14) Step of calculating the center point of the sample pixel S16 calculates the Mahalanobis distance between the sample pixel and the center point to obtain the sample 21 200834476. The distance value step S18 defines a threshold value according to the sample distance value. Step S20 takes one of the image pixels of the image plane as the determination pixel. Step S22 calculates the judgment pixel and the center. The pixel distance of the point is obtained by a pixel distance step S24 to compare the pixel distance and the sample distance, and when the pixel distance is smaller than the sample distance, it is determined that the image mask has a fire pixel. Step S50 Sample Analysis Step S52 Image Capture Step S54 Fire pixel determination step S55 fire 昼 surface counting step S56 fire 昼 surface count up to 30 steps S58 issued an alarm step S59 If the image 昼 surface is not a fire 昼 surface, the fire 昼 surface count returns to zero 22

Claims (1)

200834476 X. Patent application scope: 1. A fire image detection method, which compares a fire image sample with at least one captured image frame to determine that the captured image mask has a mask The fire image detection method includes: converting each sample pixel of the fire image sample into a color space, the color model having a first coordinate and a second coordinate, wherein the sample pixel is Corresponding to the first coordinate and the second coordinate having a first coordinate value and a second coordinate value; calculating a first coordinate value of the sample pixel and a center point of the second coordinate value; calculating the fire image sample a plurality of sample distance values obtained by each of the sample pixels and a Mahalanobis Distance of the center point; a threshold value is defined according to the sample distance values; and a plurality of images of the captured image plane are taken One of the pixels is a judgment 23 200834476 pixel; calculating the Mahalanobis distance of the judgment pixel from the center point to obtain a pixel distance; and comparing the pixel distance The sample distance, and determines the threshold at the time of a small distance to the pixel of the captured image frame pixel having the fire. 2. The fire image detecting method according to claim 1, wherein the color model is a YIQ domain, and the first coordinate is a Luminance coordinate, the second coordinate It is the color difference in-phase component (Inphase) coordinate. 3. The fire image detecting method according to claim 2, wherein the first coordinate value of the sample pixel is a brightness value, and the second sample value of the sample pixel is a color difference Component value. 4. The fire image detection method according to claim 1, wherein the Mahalanobis distance is (xz.) = ^/(x factory μ)τ Σ -1 (X factory μ), wherein Χζ· For the first coordinate value and the second coordinate value, the reed is the 24th 200834476-coordinate value and the second coordinate value of the center point, [-1 is the joint group covariate matrix, Α (χζ·) It is the Markov distance. 5. The fire image detecting method according to claim 1, wherein the threshold value is greater than 95% of the sample distance values. 6. The fire image_method of claim 1, wherein the threshold value is the median of the sample distance values. 7. If applying for the fire image_method described in item ,, the threshold value is the maximum value of the sample distance values. 8. The method of fire image grading according to item (i) of claim 5, wherein the door I value is an average of the sample distance values. 9. The fire image detecting method according to claim 1, wherein the door pinch value is the average value of the sample distance values multiplied by a safety system of 10', such as the application for the second item The fire image detection method described above has a safety factor of 1.2. u·Applicable to the fire <c< image detection method described in item 1 of the paste range, wherein 25 200834476 is the threshold value of the sample distance value, which is three times the standard deviation of the distance value of the Ryukyu Temple sample. . 12. If you apply for a patent scope! The fire image_method of the item, wherein the plurality of images of the captured image image 1 is one of the steps of determining the pixel is to take the image of the captured image in the captured image. One of the plurality of image pixels of the present invention is the fire image detecting method, wherein the fire image detecting method according to the first aspect of the invention, wherein the determining the pixel and the center point of the determining point to obtain a pixel The step of distance is to first convert the judgment pixel into the first coordinate value and the second coordinate value, and then calculate the Mahalanobis distance between the determination pixel and the center point. 4. As described in claim 1 The fire image detecting method is a fire image sample. The fire image detecting method according to the first aspect of the invention, wherein the fire image sample is a smoke image sample. I6. The method for detecting fire images as described in claim 1 wherein 26 200834476 the fire image sample is a fire image sample and a smoke image sample, the center point system Is containing - ^ fire point - the center point of the smoke, which contains sample distance - sample fire and Lin - Series tobacco samples, the pixel-based distance Included as a fire pixel distance and a smoke pixel distance, the threshold value includes a fire threshold value and a smoke threshold value, when the fire pixel distance is less than the fire threshold value and the smoke pixel distance is less than the smoke threshold value , judging that the captured image mask has a fire pixel. · 17· The method for detecting fire images as described in item 1 of the patent application, further includes the step of: sending a warning when 30 consecutive images of the captured image have the fire image. The fire image detecting method according to claim 1, wherein the step of calculating the Mahalanobis distance of the determining pixel and the center point to obtain a pixel distance further comprises converting the determining pixel into the color model. The steps. 19. The fire image detecting method according to claim 1, wherein the center point is the sample pixel of the sample. The number of tubes of the set value is about 27 200834476 and the second coordinate value. The arithmetic mean. 28