EP1323133A1

EP1323133A1 - Method and device for characterising or controlling temporal fluctuation zones of a scene

Info

Publication number: EP1323133A1
Application number: EP01972169A
Authority: EP
Inventors: Jacques Dugue; Lynda Mouloudj; Isabelle Hibon; Richard Soula
Original assignee: Air Liquide SA; LAir Liquide SA a Directoire et Conseil de Surveillance pour lEtude et lExploitation des Procedes Georges Claude
Current assignee: LAir Liquide SA pour lEtude et lExploitation des Procedes Georges Claude
Priority date: 2000-09-21
Filing date: 2001-09-19
Publication date: 2003-07-02
Also published as: US7013022B2; FR2814265A1; US20020051579A1; FR2814265B1; TWI231352B; AU2001291958A1; JP2004510236A; WO2002025590A1

Abstract

The invention concerns a method for processing images of at least a flame or a scene in a furnace, characterised in that it comprises at least a moving statistical processing of images of the flame sensed during a moving time interval, to eliminate therefrom fast fluctuations.

Description

Method and device for characterizing or controlling zones of temporal fluctuations of a scene Technical field and prior art

The present invention relates to the field of characterization and / or control of zones of temporal fluctuations of a scene.

More specifically, it relates to a method and a device for characterizing and / or controlling zones of temporal fluctuation of a scene, implementing an image processing system.

It applies to the characterization and / or control of flames, for example in an oven, in particular an industrial oven, or in any other type of environment.

The invention can also be applied to other characteristics of industrial ovens.

In general, the invention makes it possible to characterize, from a video signal coming from a camera in a fixed position, the zones of temporal fluctuations of a flame or of a scene in an oven. The invention makes it possible inter alia to distinguish and / or separate the zones of temporal fluctuations from the static zones.

The control of the thermal state of an industrial furnace is usually carried out by the use of a small number of sensors. The commonly used sensors are gas sampling systems which allow the characterization of the composition of the fumes and / or thermocouples which provide a measurement of the local temperature of the walls of the furnace or of the load, and / or sensors which carry out a measurement. spectroscopic combustion along an optical axis (called "line-of-sight") for the purpose of monitoring or controlling flames (safety).

However, the temperature sensors present in an oven, due to their low number, cannot always provide information on the drift of the flame characteristics over time. Failure to detect a flame that does not conform to its geometry, length or position may result in premature wear of the refractory walls of the oven, degraded quality of the product produced, and the emission of pollutants beyond the limits. environmental standards.

Video cameras are sometimes used on industrial ovens to provide operators with a view of the interior of the oven. The quality of visual information is however limited by the character strongly fluctuating flame, due to turbulence, and by the subjective nature of the interpretation.

Very recently, commercial computer systems have appeared which offer continuous flame monitoring by analyzing video images. These flame characterization systems, however, do not address the problems mentioned above.

Patent US Pat. Again, this type of device does not provide information on the drift of the characteristics of the flames over time.

US Pat. No. 5,971,747 describes an automated combustion control system, which implements video cameras, image processing by a neural network, and a fuzzy logic control system. It is combined with other types of sensors, such as photodetectors, temperature sensors, or pressure sensors. Such a system is very complex and does not make it possible to solve the problems of wall wear, degraded quality of the product produced, and pollutant emissions linked to the drift of the characteristics of the flames over time.

Finally, none of these known techniques is compatible with characterization and / or regulation of a flame, outside of an oven, for example in the open air. There are also no known techniques which can be applied to the characterization and / or control of temporal fluctuations of a scene of an industrial oven, such as for example the charge of an oven.

SUMMARY OF THE INVENTION The object of the invention is to provide a method and a device making it possible to follow the geometric characteristics of a flame, or of a scene in an industrial oven, in order to detect nonconforming operations, for example d '' a burner, or a fuel and / or oxidant supply, or operations which, in the case of an oven, can cause premature wear of the refractory walls of the oven and which, in any case, can lead to degraded quality of the developed product, and / or the emission of pollutants beyond the limits of environmental standards.

According to the invention, a “sliding” statistical processing of images of flames or images of a scene in an oven, obtained for example using cameras, is carried out using image processing. , said processing eliminating rapid fluctuations in the content of the images.

The images are taken during a sliding time interval, the duration of which can be variable, in particular due to the greater or lesser speed with which the statistical processing is carried out. According to another aspect, the invention also relates to a method for processing images of a flame or a scene in an oven, characterized in that, after having acquired n images of flame or of the scene:

(a) at least one statistical processing of the last n images is carried out

(b) a new image is acquired when the processing (a) is completed (c) step (a) is repeated.

The acquired images are stored for statistical processing, the latter taking into account only the last n images acquired or recorded or stacked. The acquisition of a new image causes the elimination, from the memory or of the stack in which the images are memorized, of the most anciently acquired image.

Preferably, the statistical processing is carried out on a series of images such as:

1. the delay between the last image processed (the present instant) and the oldest is between 5 and 1000 seconds, preferably between 20 and 200 seconds, 2. and / or the number of images contained in the series and participant the calculation of a statistical result is greater than 5, preferably between 25 and 1000.

According to another aspect of the invention, the statistical processing is carried out on a series of images such that the delay between two successive images is between 5 seconds and 1000 seconds, preferably between 20 seconds and 200 seconds.

The statistical processing carried out can be a calculation of variance of the images over time.

It can also be a sliding average calculation of the images of the flame or of the scene taken during the sliding time interval. It can also be a processing making it possible to obtain, from each image, an image of its instantaneous fluctuations, or else still a treatment making it possible to obtain the spatial envelope of the fluctuations of the flame or of the scene.

It is also possible to select the points of the image obtained whose intensities are above a certain threshold. This method can also include a step of extracting the contour of the spatial envelope from the flame fluctuations or from the scene.

Likewise, it may also include a step of determining a rectangle which contains or which encompasses the outline of the flame or zones of temporal fluctuation of the scene, and / or a calculation of the center of gravity and / or of the area and / or perimeter of this outline.

An image processing method according to the invention can be combined with a step of regulating a physical parameter of a flame, or of a combustion or of an oven in which the combustion or the scene takes place, or one or more burners. More precisely, we can distinguish between the parameters to be regulated (geometric characteristics of the flame) and the parameters on which we act.

The invention also relates to a device for implementing a method according to the invention, in particular as described above.

Thus, the invention also relates to a device for characterizing images of a flame or a scene in an oven, comprising means for carrying out a sliding statistical processing of the images. As explained above, such treatment makes it possible to eliminate the rapid fluctuations of the flames or of the scene.

The invention also relates to a device for processing images of a flame or a scene in an oven, comprising:

- means for memorizing n images of flame or of the scene, acquired chronologically, - means for carrying out at least one statistical processing of the n last images,

- Means for storing an additional image, when the statistical processing is finished.

Brief description of the figures The characteristics and advantages of the invention will appear better in the light of the description which follows. This description relates to the exemplary embodiments, given by way of explanation and without limitation, with reference to the appended drawings in which: FIGS. 1A, 1B and 2 represent time diagrams of image acquisition according to the invention,

FIG. 3 is a flow diagram representing one aspect of a statistical processing according to the invention,

FIGS. 4, 5 and 6 respectively representing an instantaneous image of a flame, a sliding average of several instantaneous images and an envelope of instantaneous flame,

FIGS. 7, 8, 10, 11 represent various images obtained by image processing according to the invention,

FIG. 9 is a histogram of the intensity levels of an image,

FIGS. 12 and 14 are schematic representations of devices according to the invention,

- Figure 13 shows various components of a computer system.

Detailed description of embodiments of the invention

According to the invention, a "sliding" statistical processing is applied to the images.

This processing is applied to n successive images as illustrated in FIG. 1A. A sliding interval of duration Tg is defined, in which n images lmg (i), i = 1, ..., n are recorded and memorized, at times 1, 2, ..., n.

These n images are processed, according to a statistical treatment such as one of those described below. At time n + 1, a new image lmg (n + 1) is memorized, and the statistical processing is applied to the images lmg (2), ..., Img (n + 1).

The image lmg (1) is deleted from the memory or from the stack of images to be processed. The n images are located in an interval of duration Tg, which moves with a step of duration δt, predetermined.

However, as illustrated in FIG. 1B, if, at time n + 1, the statistical processing of the n images is not completed, the image acquisition lmg (n + 1) does not take place.

If the statistical processing of the images lmg (1), ..., Img (n) is finished at the instant n + 2, the image lmg (n + 2) is recorded and a statistical processing is applied to the n images! mg (2), ..., Img (n-1), lmg (n), lmg (n + 2). The sliding time interval is then variable (Tg1 ≠ T'g2).

This is reflected in the flowchart in Figure 3.

The last n images are processed statistically (step 10). It is then checked whether the statistical processing has ended (step 12).

The acquisition of a new image (step 14) takes place only if the last n previously acquired or stored images have been processed.

According to yet another aspect, it is sometimes desirable to impose a minimum time interval between the acquisition of two consecutive images. This may be the case if the computer performs the calculations quickly relative to the desired time window.

Preferably, the calculations are done by recurrence: when a new image is captured, the algorithm does not need to recalculate all the images in the stack. It is enough to take into account (to add) the contribution of the new image, and to withdraw the contribution of the oldest image which will be overwritten.

In certain cases, for example in the case of oscillatory combustion, it may be advantageous to select images in phase with the combustion cycle. This is the case in FIG. 2, where the images are recorded at times t1, t2, t3, t4, corresponding for example to determined phases of a combustion process, for example to a determined state of opening of a fuel or oxidant supply valve.

In the case of FIG. 2, the upper part represents the time evolution of the flow rate Q of a fuel injected into a burner. This evolution is periodic (here: sinusoidal) and the image acquisition takes place only when the flow rate Q is at a predetermined phase of his cycle. Other examples may relate to any other type of variation or periodic phenomenon of an oscillatory combustion, the image acquisition being synchronized with this variation or this periodic phenomenon or being carried out only for a predetermined phase of this variation or periodic phenomenon.

Here again, it is preferable to wait for the end of the statistical calculation performed on the last n images before acquiring a new image, as already explained above in connection with FIGS. 1B and 3.

The statistical processing implemented according to the invention is sliding, which means that it is carried out on a stack of n sliding images: it is therefore not frozen over time. When a new image is captured (for example: image n + 1 in figure 1A, or image n + 2 in figure 1B), it overwrites the oldest image (image 1) in the stack . Thus, the memory allocation for the algorithm always corresponds to the number of images of the stack on which the statistical processing is applied.

The number n of images is chosen by the user when the calculation is initialized.

A number greater than 5, for example between 5 or 10 and 1000, for example equal to 20 or 25 or between 25 and 1000, or still between 20 or 25 and 200, or even greater than 50, makes it possible to obtain a sufficient smoothing effect of video noise on all n images. All fluctuations in the image are considered noise to the eye. Therefore, fluctuations in the flame or the scene viewed in an oven, as well as the background of the image, will also be smoothed.

The time interval Tg used to select the images is preferably between 10 s and 1000 s. A duration of approximately 10 s is well suited to the nature of the turbulence that can appear in a flame or in an oven, the turbulence having durations of less than 10 s. The upper limit of the interval (approximately 1000 s, or more than ten minutes) is chosen so that the information resulting from a statistical processing of images acquired during this period still has meaning for the observer, in relation to the combustion process in progress, or does not arrive to him with too much delay in relation to this process. It follows from the explanations given above in connection with FIGS. 1B and 3 that this time interval Tg can be variable. Thus, in Figure 1 B: T'g ₂ is different from Tg-i. In general, the length of this interval is a function of the number n of images to be acquired for statistical processing, as well as the speed of calculation of the computer which performs the statistical processing. A user can define, before any statistical calculation is carried out, an (approximate) duration for Tg. However, an actual acquisition does not take place over a duration strictly equal to Tg, the acquisition duration being able to be greater or slightly higher when a statistical calculation is not completed at the time when a new acquisition should be made.

FIG. 4 represents an example of an instant image Img.

We can see the walls 20 of the oven and the burner 22, with its different fuel injection and oxidant orifices.

A first example of statistical processing on n images is a variance calculation on these n images. An image is therefore obtained of which each pixel or zone is the result of the variance calculation for the pixel or the corresponding zone on all the images. High intensity values in this image correspond to the areas where the intensity fluctuations are high, and low values correspond to the areas where the fluctuations are small.

Variance is defined as the difference between the average image of a sequence of images and a snapshot. For a stack of n images, the variance (denoted V) is expressed for example by the formula:

with x the average image of the stack, and x the instantaneous image.

The Variance tends towards zero if the images are all identical, that is to say if there is no movement from one image to another. Indeed, there is then no difference between the averaged image and an image Instant. In the present invention, the variance informs about the fluctuations of the flame.

The standard deviation image is defined as follows in the article "Average Centerline Temperature of a Buoyant Pool Fire Obtained by Image processing of video recordings", L. Audouin, G. Kolb, et al (1995), Laboratoire de combustion et de detonics (University of Poitiers):

r

where I is the gray level of each pixel for each snapshot. The variance is the square of the standard deviation.

A second example of statistical processing on n images, which can be applied in the context of the present invention, is a sliding average processing (or "running average" in English).

The sliding average calculation is used for the algorithms of the envelope of the flame fluctuations or of a scene in an oven and can also be used for the calculations of color ratios. It is for example an operation of summation of matrices, point by point, carried out for example on images at scale ^Λ A (384 * 288 pixels). The mathematical expression of the algorithm is as follows:

^X t + 1 - ^X t ⁺ where:

- 1 + 1 is the time of acquisition of the image i + 1 - 1 is the time of acquisition of the image i - Xj ₊ ι represents the image i + 1, newly acquired

- xi represents image 1, that is to say the oldest image in the stack; this image will be overwritten by the new image that will enter the stack.

This algorithm is recurrent, in the sense that it does not require recalculating the average from all the initial images; it only requires the addition of the quantity (X _{i + 1} -Xj) / N after each new acquisition. Figure 5 shows an average over 60 images. This figure shows the static areas 20 of the image, which are the refractory walls. Reference 22 designates the burner output and reference 24 the smoothed image of the flame. Typically, on an image taken in an industrial oven (for example: image of a flame or of a bath in an oven), the static zones are the refractory walls and the fluctuating zones are the turbulent flames and / or, possibly, the charge (glass bath, molten metal ...). The sliding average processing gives a smoothing effect on the information and therefore tends to eliminate the fluctuation zones of the flame or of the scene viewed.

The result image obtained is of good quality, independently of the processing time, of the noise due to the video acquisition electronics.

This result image is used to more easily determine the threshold to be chosen for the following calculations. Indeed, we can determine an intensity value or the maximum intensity value on the walls

(zones 20 of FIG. 5). The threshold is then selected at this wall intensity value.

A third processing example or processing algorithm makes it possible to follow the fluctuations of the flame image by image.

This example makes it possible in fact to obtain an intermediate result for subsequent calculations.

According to this processing, the image of the moving average Avg is calculated and is then subtracted from each of the instant images of the stack. Then, the absolute value image of each of the result images is binarized according to a threshold chosen by the user.

Finally, it is checked whether the statistical processing is finished. If not, it is completed by returning to the previous steps. If so, a new image is acquired. The image of rank 1 is then overwritten or eliminated from the stack and replaced by the image of rank 2. Similarly, the image of rank i + 1 replaces the image of rank i, and this for all i between 1 and N. The new average is calculated on the new n last images thus obtained.

Subtracting the average from each snapshot removes the static envelope from the flame. Only its fluctuations remain instant. A simple mathematical expression of the algorithm is as follows:

Bin (Abs (lm (i) - Avg), threshold) Where: - Im (i) is the instantaneous image taken at time i,

- Avg is the average over the last n images,

- Abs is the absolute value calculation operation,

- Bin is the binarization operation carried out with respect to an intensity threshold chosen by the user (the pixels of Abs (lm (i) -Avg) whose intensity exceeds, or is equal to, this value is set at 1, those below are set to 0).

An example of the image obtained (also called: instantaneous flame envelope) is shown in FIG. 6.

This image helps to detect punctually abnormal states of the flame (drift or variation of its geometry).

This third algorithm makes it possible to highlight the zones where the light intensity fluctuates over time. It also removes the areas where the light intensity is constant, and underlines the edges of the flame. A fourth and a fifth algorithm make it possible to detect the envelope of the fluctuations of the flame.

The fourth algorithm implements the third processing example above. It uses the moving average as a reference image to subtract the background noise from the resulting image. For each image in the sequence, the absolute value of the difference between the image and the averaged image of the series is calculated. This result image is then binarized with respect to an arbitrary threshold, then averaged with the other images in the series.

It is then checked whether the statistical processing has ended. If not, it is completed by returning to the previous steps. If so, a new image is acquired. The image of rank 1 is then overwritten or eliminated from the stack and replaced by the image of rank 2. Similarly, the image i + 1 replaces the image i, and this for any rank i between 1 and n. The new average is calculated. The expression of the algorithm for calculating the flame envelope can be summarized by the following expression: Flame envelope 1 = Avg (Bin (Abs (lmg (i) - Avg), threshold))

Where: - Avg is the running average, as already explained above (second example of treatment).

- Bin is the binarization operation according to the threshold defined by the user (the pixels of Abs (lmg (i) -Avg) whose intensity exceeds, or is equal to, this value is set to 1, those which are below are set to 0), - Im (i) represents each instant image in the image stack.

The choice of the threshold depends on the gray level values of the starting image (average image of the n last images). These values are used to separate the background noise from the foreground flame. The threshold therefore makes it possible to identify the zones of fluctuations whose amplitudes exceed this value. This user-defined threshold eliminates all low intensity fluctuations caused by video noise in static areas. Thus, only the zones of intensity fluctuations greater than the threshold are visible on the result image.

An example of an image obtained by this fourth algorithm is given in FIG. 7.

The fifth algorithm consists in calculating the absolute value of the difference between a snapshot and the one that precedes it in the stack. Subtracting two consecutive images allows you to highlight what has changed between the two moments. The resulting image is transformed into a binary image according to an arbitrary threshold. It is then averaged (on a sliding average) with the other images in the stack.

It is then checked whether the statistical processing has ended. If not, it is completed by returning to the previous steps. If so, a new image is acquired. The image of rank 1 is then overwritten or eliminated from the stack and replaced by the image of rank 2. Similarly, the image i + 1 replaces the image i, for any value of i. The new average is calculated.

The fifth algorithm is expressed mathematically as follows: Flame envelope 2 = Avg (Bin (Abs (lmg (i) - lmg (i-1)), threshold)) Where:

- Avg is the sliding average, - Bin is the binarization operation according to an arbitrary threshold (the pixels of Abs (lmg (i) -lmg (i-1)) whose intensity exceeds, or is equal to, this value is set to 1, those which are below are set to 0),

- Im (i) represents each instant image of the acquisition stack, - lm (i) and lm (i-1) are two successive images of the acquisition stack.

Figure 8 shows the result of a "flame envelope 2" algorithm.

The images in FIGS. 7 and 8 are examples obtained by executing the algorithms on stacks of 60 images. In both cases, the static oven walls have disappeared from the picture. They are black and the flame outline is smoothed.

For the fourth algorithm, the result image is no longer sensitive to the fastest variations of the flame.

The difference between the fourth and fifth algorithms is that the fourth is influenced by the contribution of the last n images of the stack while the fifth is a more instantaneous result.

By subtracting the average, the fourth algorithm highlights the constant part of the flame (in black at the heart of the flame in Figure 7). Therefore, the fluctuating edges of the flame are also better highlighted in this result image. The information of the fifth algorithm is more instantaneous.

A sixth example of calculation or algorithm is a calculation of "instant intermittence" prior to the seventh algorithm. Such a calculation, applied to the initial images, depends on another threshold chosen by the user of the software. Each pixel exceeding the chosen value is set to 1 (white on the image) and the pixels below the predetermined threshold are set to 0 (black on the image).

This algorithm is written mathematically as follows: Instant intermittence = Bin (lm (i), threshold) An instant intermittence image is a binary image which informs about the probability of exceeding a threshold set by the user (p = 0% = black or p = 100% = white). Consequently, the chosen threshold determines the presence of flame in the image. The binary image, whose pixel values are 0 or 1, will be white at the flame location and black if no pixel exceeds the threshold value. The histogram of the grayscale values of a flame image can help the user to choose the threshold. The threshold value is generally located above the maximum intensity of the walls. Figure 9 shows the typical histogram of an image taken in an oven. One can carry out (seventh example of calculation or algorithm) a calculation of "average intermittence", which is the sliding average in time of the instantaneous intermittences. The pixels of the image therefore vary from 0 to 255, and their gray levels (or intensities) represent the probability of exceeding the chosen threshold, during the time interval of the sliding window.

The image of the probability of exceeding a brightness threshold greater than the brightness of the walls therefore makes it possible to show the presence and effective position of the flame in the image as shown in FIG. 10. The mathematical expression of this algorithm is as follows:

Avg (Bin (lm (i)), threshold), Where:

- Avg is the sliding average,

- Bin is the binarization operation of the instantaneous image Im (i) according to the arbitrary threshold: a pixel of lm (i) is set to 0 (or to 1) if its intensity is lower (or higher) than the threshold.

The use of a threshold according to one of the third to seventh algorithms is advantageous for erasing the noise of the image and the brightness of the walls and extracting the flame contour. This type of algorithm has certain limits. Indeed, one can thus characterize the flame only if it is brighter than the bottom. However, this is not always the case on industrial sites, in particular because of the low luminosity of certain flames compared to the radiation from the refractory walls. The image of the result is dependent on the choice of the threshold, but this is compensated for by the systematic aspect of this dependence which allows a comparison of the image results over time.

The fourth, fifth and seventh algorithms implement the AVG (BIN (, threshold)) function, the argument being, respectively,

Abs (Img (i) -Avg) (fourth algorithm), Abs (lmg (i) -lmg (i-1)) (fifth algorithm) and Abs (img (i)) (seventh algorithm). Consequently, according to another definition of the invention, a statistical processing implemented comprises:

- the calculation of a binary image, compared to a threshold, from an argument image, a pixel of this argument image being set to 0 (or 1) by the binarization operation, if its intensity is lower (or greater or equal) to the threshold,

- a calculation of the average of the binary images thus obtained. The argument image can be for example one of the three types of images indicated above Abs (lmg (i) -Avg), Abs (lmg (i) -lmg (i-1)) or Abs (lmg ( i)).

Beyond the simple use as display of an image easier to interpret by an operator, the envelope of the fluctuations of the flame, obtained by one of the fourth and fifth algorithms or the images of instantaneous or average intermittence ( sixth and seventh algorithms), can (can) be used to extract quantitative geometric parameters such as perimeter, area, flame length. It is possible to apply an object detection method (here the flame) by image processing, for example by contour extraction, or even "contour segmentation". According to an example of extracting contours, an image is first binarized with respect to a gray level chosen by the user. For example, each pixel of the image is assigned the value 1 or 0 depending on whether it is considered to be part of the flame, or not. In a second step, the image is expanded by one pixel wider. Then the original image is subtracted from the expanded image by one pixel. The result of this subtraction is one or more continuous contour (s), one pixel thick. Finally, this or these contours are superimposed on the result image (for example "flame envelope 1" or "flame envelope 2"), by adding the two images. Figure 11 shows an example of contour extraction with a rectangle enclosing the flame. The coordinates of the enclosing rectangle (x _m j _n , x _m ax_ Ymin, Ymax) and / or one or more other parameters such as its center of gravity, or the area of the contour or its perimeter can be calculated and displayed. The results are sent to files or memory areas to allow archiving and monitoring of the characteristics of the flame over time. This analysis gives a good number of quantitative geometric parameters. They can be dynamically linked to advanced control systems such as neural networks. They can thus be used as additional inputs for online flame control.

The monitoring of these flame contour parameters can be advantageously used to maintain an optimum adjustment of one or more type (s) of parameters of an oven and / or of combustion and / or of one or more burners, for example one or more of the following parameters:

1. The pressure of the spray fluid.

In the case of a burner operating with liquid fuel, the flame envelope parameters, and in particular the position of the flame root, can be used to regulate the spraying conditions, and in particular the flow rate and / or pressure. Too low a pressure usually results in a flame that is too long, with a flame root farther from the injector.

2. The degree of flame staging.

For burners that divert part of the fuel or the oxidant to a secondary injector, the flame outline can be used to regulate the degree of staging (therefore the proportion of fuel or oxidant to be directed towards the secondary injector) and optimize the length and volume of the flame. We can seek in particular to avoid situations where the flame is too close to the thermal load (glass bath, metallurgical products) or refractory walls. Regulation of the degree of staging can also be used to minimize pollutant emissions.

3. Fuel and oxidant flow rates.

The fuel and oxidant flow rates, as well as the oxidant / fuel flow rate ratio can be used to maintain a correct flame envelope. Indeed, an oxidant / fuel flow ratio lower than the stoichiometric ratio generally results in an excessively long flame, and a total flow - {oxidant + fuel) too low compared to the nominal power of the burner can cause the flame to rise towards the roof of an oven.

4. The power and the momentum of the neighboring burners. If several burners are used, the image processing system can be used to simultaneously follow the envelopes of several flames, and will make it possible to diagnose or identify undesirable interactions between the flames of neighboring burners. The information from the flame envelopes can be used to optimize the position, the injection modes and the amount of movement of the fluids (mass flow and speed of the fluid) of each burner so as to avoid these undesirable interactions (the amount of movement of a fluid is equal to the product of the mass flow by the speed of this fluid). These conditions concern in particular glass furnaces as well as certain metallurgical furnaces (reheating furnaces).

5. The fraction of waste introduced into the burner.

In the case of combustions where waste is co-incinerated with conventional fuels, the image processing system can be used to control the fraction of co-incinerated waste to a shape and / or to a flame root position. This could for example be the case of cement kilns, where it is desirable to maximize the fraction of energy provided by waste, while respecting satisfactory combustion characteristics (stable flame, flame root immediately downstream of the burner).

6. A fraction of total oxidant introduced by the burner.

In the case of burners which combine oxidants of different oxygen concentrations (for example air and oxygen, recycled fumes and oxygen, etc.), the flame contour can be used to regulate the ratio between the two oxidants so as to maintain the flame length in an acceptable range. Indeed, the increase in the overall oxygen content in the oxidant generally results in a shortening of the length of the flame.

7. The oven pressure. The presence of air inlets near a burner can have significant consequences on the direction and shape of the flame. The information on the flame envelope, possibly in combination with that from other sensors, can therefore be controlled by a parameter which controls the air inlets of a furnace. This parameter could be the position of a valve in the flue gas exhaust to act on the pressure inside the oven. It is also possible to act on the air inlets by a maintenance action aimed at improving the seal around the burner. Variations in the position of the flame envelope can indeed be a sign of the presence of parasitic air inlets in the oven.

8. The frequency of oscillation of the oxidant and fuel supplies of a burner. In the case of the use of a valve allowing an oscillatory combustion, the acquisition of images can be synchronized in phase with the valve, and the image analysis can allow a statistical treatment on the flame or the envelope of the flame for different phases of the oscillations of the fuel / oxidant mixture. Image processing according to the invention makes it possible to verify that, for each phase of the oscillation cycle, the flame envelope maintains acceptable characteristics. Control by video analysis, for example in combination with other temperature and composition sensors, of the fumes, allows optimization of the frequency and / or amplitude of the oscillations so as to minimize the emission of pollutants while keeping a flame envelope compatible with the process.

FIG. 12 shows an example of a device for implementing the invention in an industrial oven 40.

This example is given for viewing a flame. It also applies to the observation of a load in an oven.

A burner 42 is shown diagrammatically, as is a flame 44.

Image acquisition means, such as one or more camera (s) 46, make it possible to acquire images of the flame 44. These images are processed by a device or card 48 for digitizing images. Video cameras used in industrial ovens can operate in the visible, ultraviolet or infrared. To increase the contrast between the flame and the refractory walls, these cameras can be equipped with an interferometric filter (in the ultraviolet: filter centered around 310 nm to highlight the emission of the OH radical; in the visible: filter centered around 431 nm for the CH radical, or 516 nm for the C2 radical, or 589 nm for the emission of sodium; the bandwidth of the filters is between 10 and 20 nm).

The digitized images are transmitted to computer means 50, essentially comprising a central unit 60, display and display means 69, and control peripherals such as a keyboard 72 and a mouse 61. Other selection means of a zone or of a field of a page displayed on the screen 69 can also be used, for example any means allowing a selection to be made by touch on the screen.

In the case of oscillatory combustion, additional data is introduced into the computer system 50: it is a digital signal representative of the oscillatory periodic signal.

As illustrated in FIG. 13, the central unit 60 itself comprises a microprocessor 62, a set 64 of ROM and RAM memory, a hard disk 66, which also has an information storage function, all of these elements being coupled to a bus 68.

Screen 69 makes it possible to view one or more of the raw images (before statistical processing) or of the images obtained after statistical processing. In FIG. 12, the screen 69 is represented with an instantaneous image 69-1, an average intermittency image 69-2, a flame envelope image 69-3 and a contour image 69-4.

The instructions for implementing a statistical processing according to the invention are stored in the means 64, 66 for storing the computer system.

Means, for example a menu and a cursor moved with the mouse, allow a user to select the statistical processing to be carried out (for example one of the first to seventh processing or algorithms which have been explained above. ). He can also choose to carry out several of these treatments in parallel. Identical or the same type of means can also offer the user the possibility of choosing the number n of images to be acquired in order to carry out a sliding statistical processing, and / or the duration, exact or approximate, of a sliding time interval. Identical or same type means can also offer the possibility of selecting one or more threshold values, for the implementation of one or the other of the algorithms described above.

Identical or same type means can also offer the possibility of selecting one or more argument images with a view to statistical processing implementing the AVG (BIN (, threshold)) function mentioned above.

The raw images acquired using the camera 46 and the digitization card 48 are stored in a memory area of the central processing unit 60. In fact, a set or a stack of the last n images is stored in this memory area. acquired, or images acquired during the duration of the selected sliding interval.

Also stored in a memory area are the last n images obtained by sliding average Avg, or a stack of these last n average images, or other stacks of images which evolve during the acquisition (for example the stack of results of

Abs (lm (i) -lm (i-1)).

The display method can also indicate to the operator, the light intensity corresponding to a portion or an area of an image displayed on the screen 69. This function is implemented by means of selecting a portion. or an area of the image, for example using the cursor, and by means of display, on the image, for example in a determined field thereof, of the intensity of the selected area .

The user can then set a threshold value in relation to such information, for example by selecting a specific field on the screen.

In outline display mode (figure 11 and image 69-4 in figure 12), the quantitative values of the coordinates of the outline frame are also displayed, and possibly the calculated values such as the center of gravity, and / or the contour area and / or the perimeter of this contour. The instructions of the programs for implementing a method according to the invention are stored in a memory area of the computer system 50. These instructions are for example installed from a medium which can be read by this system, and on which they are recorded. . Such a medium can be for example a hard disk, a ROM read-only memory, a compact optical disk, a dynamic random access memory DRAM or any other type of RAM memory, a magnetic or optical storage element, registers or other volatile memories. and / or non-volatile. The device can be used to view instant images or those resulting from statistical processing. This information is already very useful for monitoring and understanding a combustion.

From this information, an operator can optionally act on the oven or burner (s) control parameters (for example power, and / or stoichiometric ratio, etc.) in order to control one or more parameters characterizing the position and geometry of the flame or flames (in the case of several burners).

It can also be one of the parameters 1 to 8 already mentioned I c above.

As illustrated in FIG. 12, the device can also comprise means 52 for regulating parameters, for example one or more of the parameters 1 to 8 mentioned above. This regulation can be carried out, for example, from an analysis of the images obtained by statistical processing, for example an analysis using neural processing and / or control by fuzzy logic. The command 52 then makes it possible to regulate, for example, the opening of a fuel or oxidant supply valve.

According to another example of use of a device according to the invention, the images obtained by digitization can be stored on a video cassette 74 (see FIG. 14) which can then be played by a video recorder 76. After digitization, the images can be displayed on a computer system 50 as already described above. A combustion or flame analysis can thus be performed offline, in the laboratory. Each raw video image results from the combination of 3 colors or 3 channels R (red), V (green) and B (blue). It may be advantageous, in some cases, to retain only one channel. For example, in some cases, the R channel is highly saturated, the B channel has a low contribution and the V channel is best "balanced". We only select channel V.

For each type of image (instantaneous or obtained by statistical processing), the device can therefore include means for selecting a display of the images in only one of the colors R, G, B, or in two of these colors. These means (for example a menu in which the user selects one or more fields with a cursor) also make it possible to select, for each given type of image, a representation of the ratio of two of these colors in the image.

The images are coded on 8 bits (therefore on 256 intensity levels).

In the case of the above algorithms for which the Avg function is applied to a binarized image, each pixel is averaged with the corresponding pixels of the other images. The result is, for each pixel, an intensity value between 0 and 1, which is then converted back to full scale (on 256 intensity levels) by multiplication by 255.

All the image processing indicated in this description and which involve the choice of a threshold are, because of this threshold, arbitrary or biased. But this arbitrary character is constant over time, it is the same for all the n images of the sliding interval

Tg or for all the last n images.

Image processing according to the invention is much lighter and more demanding in terms of computing capacity than the system described in US Pat. No. 5,971,747, where neural processing is applied to each image. According to the present invention, a sliding statistical processing is applied to the images, and a neural processing as described in US Pat. No. 5,971,747 is not necessary. Such neural processing only intervenes in a possible regulatory loop, such as loop 52 described above (FIG. 12). The invention applies to viewing and controlling flames or combustion in an oven, but also in any type of other industrial environment, including in the open air.

The invention and the statistical processing described also make it possible to characterize fluctuations in a scene in an oven, for example of a charge present in the oven (clods floating on the surface of a glass bath, line showing the limit of presence of unfounded material in a melting furnace, spatial envelope of the billet trajectory in metallurgical furnaces, etc.). Any of the algorithms described above can then be applied, with the same advantages as what has been described for the case of a flame. In particular, it is possible to apply a contour extraction function to the fluctuating zone of the charge in the furnace, to deduce therefrom geometric parameters such as those already mentioned above (perimeter of the contour, and / or center of gravity, and / or area of the contour), and possibly to regulate the bath (its temperature or its supply under load) or the trajectory of the billets.

The invention makes it possible to view, in an oven, any element of a fluctuating nature over time or of a brightness different from the brightness of the environment.

Claims

1. Method for processing images of at least one flame or of a scene in an oven, characterized in that it comprises at least one sliding statistical processing of images of the flame taken during a sliding time interval , to eliminate rapid fluctuations.

2. Method according to claim 1, the time interval being of variable duration.

3. Method for processing images of at least one flame or of a scene in an oven, characterized in that, after having acquired n images of flame or of the scene:

(a) at least one statistical processing of the last n images is carried out (b) a new image is acquired when the processing (a) is completed (c) step (a) is repeated.

4. Method according to one of claims 1 to 3, the statistical processing being carried out on a series of images such as: - the delay between two successive images processed is between 5 seconds and 1000 seconds.

5. Method according to claim 4, the delay between two successive images processed being between 20 seconds and 200 seconds.

6. Method according to one of the preceding claims, the number of images contained in a series of images processed by sliding statistical processing being greater than 10

7. Method according to the preceding claim, the number of images contained in a series of images processed by sliding statistical processing being between 25 and 1000.

8. Method according to one of the preceding claims, the statistical processing being a calculation of variance of the images over time.

9. Method according to one of claims 1 to 7, the statistical processing being a calculation of the sliding average of the images of the flame or of the scene taken during the said sliding time interval or of the last n images.

10. Method according to one of claims 1 to 7, the statistical processing making it possible to produce, from each each image, an image of its instantaneous fluctuations.

11. Method according to claim 10, the processing of the images comprising the following steps:

- the calculation, for each series of n images of the same sliding time interval, or for each series of n last images, of an average image, - the calculation, for each image of said series of n images or of n last images, the difference between the image and the averaged image of the series of n images,

the comparison of each difference image thus obtained with a predetermined threshold value, in order to obtain an image known as a binarized image,

12. Method according to one of claims 1 to 7, the statistical processing being a processing making it possible to obtain the spatial envelope of the flame or scene fluctuations.

13. The method of claim 12, the processing of the images comprising the following steps:

- the calculation, for each series of n images of the same sliding time interval, or for each series of n last images, of an average image, - the calculation, for each image of said series of n images or of n last images, the difference between the image and the averaged image of the series,

the comparison of each of the n difference images thus obtained with a predetermined threshold value, in order to obtain an image called binarized image,

- the calculation of an average image of the n binarized images thus obtained.

14. The method as claimed in claim 12, the processing of the images comprising the following steps:

- the calculation, for each image of a series of n images of the same sliding time interval, or for each image of a series of n last images, of the difference between the image and the image which precedes it in series,

- the calculation of an average image of the n binarized images thus obtained.

15. Method according to one of claims 1 to 7, the statistical processing comprising:

- the comparison of each image of a series of n images of the same sliding time interval, or of each image of a series of n last images, with a predetermined threshold, to obtain a binary image,

- the calculation of an average image of the n binary images thus obtained.

16. Method according to one of claims 12 to 15, further comprising a step of extracting or segmenting contours.

17. The method of claim 16, further comprising a step of calculating the rectangle encompassing the outline of the flame or the fluctuating area of the scene and / or the center of gravity of this outline and / or the area of this contour and / or the perimeter of this contour.

18. Method according to one of the preceding claims, the images being those of the flame of a combustion taking place in an industrial oven in operation.

19. Method according to one of the preceding claims, the images being those of a flame of oscillatory combustion.

20. Method according to the preceding claim, the processed images being images synchronized or in phase with the oscillatory combustion or with a phase of the oscillatory combustion.

21. A method of regulating physical parameters of a flame or of combustion, implementing a method of processing images of the flame according to one of claims 1 to 20, and a step of regulating a parameter. flame or combustion physics or an oven in which combustion takes place.

22. Method according to the preceding claim, the physical parameter comprising at least the length or the volume or the position of the flame, or the relative position of the flames in the case of a plurality of burners.

23. The method of claim 21 or 22, the physical parameter comprising at least the flow rate or the spray pressure of a liquid fuel, or the degree of staging of a fuel or an oxidant, or the flow of a fuel or an oxidant, or, in the case of a plurality of burners, the position and / or the mode of injection and / or the momentum of movement of at least one burner, or a fraction of waste reintroduced into a burner , or, if at least two oxidants are used, a concentration ratio between these oxidants, or the pressure of an oven in which combustion takes place.

24. Method according to one of claims 21 to 23, the combustion being of the oscillatory type the physical parameter comprising at least one oscillation frequency or an amplitude of the oscillatory supplies of oxidizer and / or fuel of a burner.

25. Method according to one of claims 1 to 17, the images being those of a charge in an oven.

26. Method according to the preceding claim, the images being those of a glass bath, or of a bath in a melting furnace, or of billet trajectories.

27. Device for processing images of at least one flame or of a scene in an oven, characterized in that it comprises means for receiving images of a flame or of a scene located in an oven, and means for carrying out at least one sliding statistical processing of these images.

28. Device for processing images of at least one flame, comprising: - means for storing n images of flame or of the scene, acquired chronologically,

means for carrying out at least one statistical processing of the n last images,

29. Processing device according to claim 27 or 28, the means for carrying out at least one statistical processing being means for carrying out a calculation of variance of the images over time, and / or a sliding average calculation of the images, and / or a processing to produce, from each image, an image of its instantaneous fluctuations, and / or a processing making it possible to obtain the spatial envelope of the flame or of the scene or of their fluctuations.

30. Device according to one of claims 27 to 29, further comprising means for selecting one or more statistical processing (s) of the images of the flame or of the scene.

31. Device according to one of claims 27 to 30, further comprising means for selecting a number of images on which one or more statistical processing (s) is to be performed.

32. Device according to one of claims 27 to 31, further comprising means for selecting a duration of a sliding time interval.

33. Device according to one of claims 27 to 32, further comprising means for selecting an area or a portion of an image of the flame or of the scene, and for displaying an intensity value in this area or portion .

34. Device according to one of claims 27 to 33, further comprising means for selecting one or more threshold value (s) for the implementation of one or more statistical processing (s) of the images of flame or scene.

35. Device according to one of claims 27 to 34, further comprising means for viewing images after statistical processing (s).

36. Device according to one of claims 27 to 35, the images being recorded by one or more analog (s) or digital video camera (s).

37. Combustion system comprising a burner (42), means for injecting at least one oxidant and at least one fuel into the burner, and a treatment device according to one of claims 27 to 36.

38. The system of claim 37, further comprising means (52) for regulating a physical parameter of the flame or of the combustion or of an oven in which the combustion takes place.

39. Industrial oven, comprising a load and a load image processing device according to one of claims 27 to 36.

40. Computer program comprising program code instructions for the execution of the steps of the method according to one of claims 1 to 26 when said program is executed on a computer.

41. A computer program product comprising program code means recorded on a medium usable in a computer, comprising computer-readable programming means for performing at least one of the steps according to at least one of claims 1 to 26.