20: Flame Visualization Enhancement by Image Processing – Optical Methods for Data Processing in Heat and Fluid Flow


Flame Visualization Enhancement by Image Processing

W B Ng, K Y Cheung, and Y Zhang


This Chapter presents flame image visualization enhancement using image capturing and processing techniques. Test cases of the enhancement of flame visualization have been demonstrated. It has also been shown that the images could be pre-processed by the manipulation of the shutter speeds (or gating times) of a camera, in order to obtain the desired aspects of flame dynamics such as the time and spatial scales.

20.1 Introduction

In the last few decades, laser technology has developed progressively which enables laser-based optical diagnostic technique to play a dominant role in combustion research. Direct photography and its potential, on the other hand, has been overlooked even though it has been applied to flame studies ever since the first camera was invented. This is due to the fact that the flame light emission is integrated along the line of sight, which leads to ambiguity in spatial resolution. As a result, quantitative analysis is difficult. In addition, the sensitivity of a film is often too low to capture the flame dynamics and the images are either blurred or too dim to be of any use. The situation has changed recently since we are in the middle of a revolution of imaging techniques, sparked by the rapid progress not only in advanced image sensing technology but also in digital image processing power. Conventional film cameras are being replaced by high-end digital cameras. The captured images can be easily uploaded in to a computer directly for enhancement and processing. Much useful physical insight and information can be gained by digital image processing.

By using proper combinations of the shutter speed, sensitivity, framing rate, and resolution of a camera, the time scales of combustion may be resolved and the corresponding flame structures may be extracted temporally and spatially. For example, using a lower shutter speed and light sensitivity of a CCD (charge-coupled device) sensor, an averaged image of the flame is obtained as long as the image is not over exposed or saturated. Within the exposure time (or gating time) the flame may be locally created or destroyed many times. What is recorded is the averaged flame pattern. If a higher shutter speed is applied only smaller time-scale flame movements are recorded and the flame image is more ‘instantaneous’. The flame structures at different time-scales are needed to gain a better understanding of the processes involved. Therefore, with the aid of advanced imaging equipment, imaging capturing can also be regarded as imaging processing.

In the following sections, the experimental set-up is presented first. Then the various examples of digital enhancement of flame visualization are given.

20.2 Experimental setup

All the flame images are obtained from an impinging burner. The experimental rig consists of a burner, a steel plate, a mixing chamber, a fuel, and air supply system. The steel plate was positioned above the burner nozzle [Fig. 20.1(a)]. The reactant jet came out of the burner nozzle and impinged on the plate. The burner was attached to an adjustable platform so that the height between the burner nozzle and the plate could be adjusted. The air and fuel (propane) flows were controlled separately by rotameters and pressure gauges. The two flows were then mixed in the swirling mixing chamber, which was connected to the burner by a long flexible pipe. The pipe was bent deliberately to create secondary flow so that the fuel and air could be further mixed. Three turbulent generators [Fig. 20.1(b)] were used to modify the flame turbulence intensity. They were perforated discs, 50 mm in diameter. Each disc was cut from an aluminium sheet, 1 mm thick. Some experiments were conducted without a turbulent generator. V-shaped flames could also be created by placing a steel rod above the nozzle exit.

20.2.1 V-shaped premixed impinging jets

Five steel rods with diameters of 5.5, 6.5, 8.0, 10.0, and 12.0 mm [Fig. 20.1(c)] were used to create V-shape flames of different characteristics. The first step of this experiment was to create a conic flame (1) with uniform turbulent flow of premixed air and propane at the nozzle exit. The steel rod was then placed on top of the nozzle at the center position, either lined up parallel to the camera lens or perpendicular to the lens. As a result, an impinging V-shape flame was formed. The plate and nozzle distances were adjusted and set for each set of experiments. The details of the different cases are shown in Table 20.1.

Fig. 20.1 Experiment apparatus

Table 20.1 Experimental conditions

A CCD-TRV224 Sony camcorder was used to record the different flame patterns and flame transitions. The camera has four imaging modes but only two modes were used. These are the normal mode and high shutter speed mode. The normal mode is operated at a shutter speed of 1/50th second, and the high shutter speed mode was at 1/4000th second. All modes have a framing rate of 25 frames per second. Images were recorded on video 8 mm cassettes.

A Kodak EktaPro HS4540 Motion Analyser was also used in this experiment. It was operated at 4500 frames per second. It can only capture images in B/W format.

20.2.2 Diffusion impinging jets

The experimental rig cited above was also modified for diffusion impinging jet studies by adding a small diameter fuel tube to the central axis of the nozzle. Pure propane fuel comes out the small fuel tube and coaxial air flows through the burner nozzle itself. As a result, an impinging diffusion flame is formed. The distance between the steel plate and the burner nozzle was fixed at 120 mm.

An Olympus E-100RS digital camera was used to capture the flame images. The camera speed can be set as high as 1/10 000th second. All captured images were initially stored in SmartMedia (memory) card in digital format before transferred to PC through a cable. The camera was positioned at an inclined angle to the steel plate. Images were taken at different shutter speeds but at a fixed framing rate of 15 frames per second. The flame structures of the obtained images were useful in resolving time-scale information. Eight shutter speed modes of 1/500th second, 1/1000th second, 1/2000th second, 1/3000th second, 1/4000th second, 1/5000th second, 1/8000th second, and 1/10 000th second were applied respectively.

20.2.3 Laser sheet tomography of premixed impinging jets

Details of the laser sheet tomography experimental setup have been reported elsewhere (2). The plate and burner exit separation was 25 mm and the mean burner exit velocity was fixed at 3.1 m/s. The turbulent generator was located 20 mm below the exit of the burner nozzle. They are 2 mm thick and have hole diameters of either 2 mm or 3 mm. The perforated discs have the same blockage ratio of 0.5. The flow was seeded with fine silicon oil droplets generated by a glass atomizer. After chemical reaction in the very thin flame front, the oil droplets are destroyed. Therefore only the reactant part of the flow field could be visualized when a two-dimensional thin laser sheet (0.6 mm) was shone through the flow field. The thin laser sheet was created by a long focal length spherical lens to focus the laser beam and a cylindrical lens shortly after to expand the beam in one plane.

A Kodak EktaPro HS4540 Motion Analyser was used for imaging. The camera was operated at 9000 frames per second. For each test the digital camera could store 2025 images, which is equivalent to 225 ms in real time. The captured images are stored in the RAM bank of the camera, which can then be transferred to VHS videotape. The images stored on a VHS tape have to be digitized. A video player is connected to a computer image grabber card. The captured video clips are saved in AVI format.

20.3 Data evaluation by image processing

In this section, the post processing of combustion database (raw images) to deliver information on flame structures, flame dynamics, flame behaviours, and time-scales of combustion process has been demonstrated. It starts with the images of LST, followed by the V-shaped premixed flame and ends with the diffusion flame. The pictures in Fig. 20.2 give an impression of the images to be evaluated.

Fig. 20.2 (a) V-shape turbulent impinging jet taken by a Kodak HS4540 (4500 fps)(b) V-shape turbulent impinging jet taken by a Sony camcorder (1/4000th second)(c) V-shape turbulent impinging jet taken by a Kodak PC600 (1/50th second)(d) LST premixed impinging jet taken by a Kodak HS4540 (9000 fps)(e) ‘Star’ shaped turbulent diffusion flame taken by an E-100RS Olympus digital camera (1/4000th second)

20.3.1 LST

Figure 20.2(d) shows a typical flame image captured by LST. The grey areas represent the reactant and the thin white line marks the wall position. There are 600 images in sequence cropped to a specified rectangle shape. Most of the images of the LST experiment have intensity faded non-uniformly, the appropriate threshold value may have to vary throughout the image. Therefore, instead of global thresholding, adaptive thresholding is used to create the binary image: 1 for the burned region and 0 for the unburned region, resulting in an instantaneous binary image. The binary images are then superimposed using the arithmetical operation PLUS [Fig. 20.3(c)] and the resultant image is proportional to the degree of combustion progress. Figure 20.3(d) shows the layers of reaction progress variables of 0.8 (the most outer layer), 0.6, 0.4, and 0.2 (the most inner layer).

Fig. 20.3(a) A typical cropped image with a 256 grey scale

Fig. 20.3(b) A binary image

Fig. 20.3(c) Resultant image of operation PLUS

Fig. 20.3(d) Reaction progress variable contours

Fig. 20.3(e) A single flame front boundary

Fig. 20.3(f) 250 flame front boundaries superimposed

By using a gradient operator [sobel operation (3) in this case], the contours of the images are extracted. The width of each contour is equivalent to 1 pixel-size. Once the flame front boundaries are detected, the flame surface density can be calculated by the method of either Veynante et al. (4) or Deschamps et al. (5).

20.3.2 V-shaped flames

The original images of the V-shaped flame captured by the high-speed camera are not suitable for direct information extraction because the flame patterns are almost invisible (Fig. 20.4 left). Histogram-equalization enhancement is applied to improve the visibility of the flame patterns. However camera-streak noise has also been highlighted as well and appeared in the processed images (Fig. 20.4 right). Therefore further image enhancement such as noise removal is required.

Fig. 20.4 (Left) An original V-shaped flame image, (right) a histogram-equalization version of V-shaped flame image

Basically a digital image is made up of fundamental spatial frequency components that, when combined, make up the form of the image. A frequency-transforming process provides a pictorial view of these spatial frequency components. It converts an image from the spatial domain of brightness to the frequency domain of frequency components and a distinct white point may appear in the frequency-transform image if a repetitive noise pattern exists in the original image. The horizontal streaks shown in existing raw images (Fig. 20.4 right) are a type of repetitive noise pattern. Therefore, a two-dimensional Fourier transform algorithm is used to convert the original image to frequency domains (Fig. 20.5 left). We multiply the frequency images by 0 in the area of the spots. This creates a band-reject filter that attenuates only the frequency of interest. An inverse Fourier transform algorithm is then applied to transform the frequency images back to the spatial domain (3). As shown in Fig. 20.5 (right), the image no longer contains the repetitive noise streaks.

Fig. 20.5 (Left) Fourier-transform image showing horizontal noise patterns as bright spots in the image, (right) inverse Fourier-transform image with periodic noise removed

The quality of all raw images has been improved by means of image processing algorithms. This process allows one to further manipulate and evaluate the post-processed images in order to extract wanted information by intelligently reducing the amount of image data as each digital image contains enormous information and often much of this information is superfluous to solve a specific problem. In the following section, data extraction techniques are elucidated. Image subtraction

It is often the case that the flames have changed little within a very short period of time. As a result, it is difficult to spot the variations between two consecutive images. Subtraction of the two adjacent images would highlight the minor images. This method is often used to detect motion. Consider the case where nothing has changed in a scene; the image resulting from subtraction of two sequential images is filled with zeros and replaced by a homogeneous colour. In contrast, if something has moved in the scene, subtraction produces a non-zero result at the location of movement.

Fig. 20.6 Subtracted images

The top left subfigure shows the difference between two consecutive frames, i.e. frame 2 is subtracted from frame 1 with a time difference of 0.222 ms. It is very hard to distinguish the differences between the two images. As the gap of the frames increases, the difference and the structure of the flame movement are getting more obvious. In between Frame 5-1 and Frame 6-1 (0.889–1.111 ms), the structure of the flame movement is in its most detailed form. As the separation continues to increase, the images start to lose their information on the structure of the flame movement. When the frames gap is 7 or greater, a false impression of the flame structure movement seems to be shown. Although it still highlights the main area of the flame movement, the details of the structures are not shown. Analysis of rod diameters on V-shaped flame behaviour

There are five different sizes of rods used in the experiments. Using the image processing techniques, it is possible to evaluate the flame separation distances and angles. The first measurement is taken from Data Set 1, which has the same variables except for the varying rod sizes, as shown in Table 20.1. For each set of data, the first fifty frames are grabbed from each data clip. All outer boundaries are tracked automatically and superimposed on to one frame, where the average of the flame position can be obtained, shown in Fig. 20.7.

Fig. 20.7 (Left) 50 tracked outer boundaries of each case of Data set 1 are superimposed on to a single image, (right) A zoom in version of the processed image

The inner boundaries tracing can provide the positions as well as the movements of the flame at a specified time interval. Next, 10 frames are grabbed from each test case instead of 50, and there is a 10 frames gap between each frame. For example 1st frame = frame no.1, 2nd frame = frame no.11, 3rd frame = frame no.21, etc. The contrast level is increased for easy boundary tracking. Again, all ten tracked boundaries are superimposed on to a single frame (Fig. 20.8). The results give the movement of the flame in a time period of 22.22 ms. In general, these images show that the bottom half of the flame stays fairly steadily. The flame starts to fluctuate more vigorously further away from the rod. Near the tip of the flame, the behaviour is very unpredictable. There is no obvious pattern. The scale of turbulence becomes greater.

Fig. 20.8(a)–(d) Image contours superimposed in a single plot

The angle of the flame separation is measured from the centre of the rod, where it is marked before the image layer is removed. Two tangential lines are drawn in the inner flame boundary from the centre point of the rod. The results of the angles with different rod diameters are tabulated in Table 20.2.

Table 20.2

Fig. 20.9 Angle measurement from contour-superimposed plot

Figure 20.10 shows the angle increases when the rod diameter increases. From all the measurements taken above, it can be concluded that the larger the rod diameter, the greater the separations between the inner flame boundaries and between the rod and the flame root.

Fig. 20.10 Plot of flame separation angle versus rod diameter Analysis of the distance between the plate and the nozzle on flame structure

This is a variable which defines the vertical distance between the nozzle of the burner and the flat metal plate hanged above. Three cases have been investigated. All the other variables are kept the same, except for the nozzle to plate distance. Using a time gap of every 5 frames, 20 frames are captured in each case. Each of these frames is processed with increased contrast. A comparison of the difference in contrast is shown in Fig. 20.11.

Fig. 20.11 (Left) An original V-shaped flame image, (right) contrast enhanced image

It is easier to highlight the region of the vortex in the inner flame boundary (circle) in the above figure. In general, when a flame is impinging on a wall, it has the following behaviour.

Fig. 20.12 Impinging V-shaped flame

Region 1 The flame has a greater momentum at the exit of the nozzle. The flame is separated by a metal rod with little fluctuation and the creation of small-scale wrinkles.

Region 2 This is the impingement region, where the flame hits the metal plate and flows along the wall. The velocity in this region is gradually decreasing due to the friction resistance of the plate and the divergence of the flow. The vortex generated in Region 1 is developing with an increase in its scale.

Region 3 This is the forward vortex region, as the flame is separated at the root at an angle. The flow of the flame is rotating and rolling along the plate, and thus vortices are formed. The mixing and the diffusion are weak in this region, although the momentum is supplied from Region 1. Region 4 is similar to Region 3 due to the symmetry of the problem. Effect of turbulence generators on V-shaped flame behaviour

There are three different turbulence generators used in the experiments for comparison.

TG0 = No turbulence generator.

TG1 = Big and relatively few holes

TG2 = Smaller holes with a higher number density of holes.

TG3 = Smallest holes with the highest number density of holes.

The subtraction technique is used to highlight the structures of the flame. As seen from Fig. 20.13, the main structure of the flame is generally situated at the center region of the flame, where the rough structure of the flame can be seen for each case. It can be seen that as for TG0, the flame fluctuates very little. When a turbulent generator is used, the flame structure is more apparent, as the generator will create more turbulence. Therefore, changes in flame movement will be more vigorous.

Fig. 20.13 Subtracted image of TG0, TG1, TG2, and TG3 at top left, top right, bottom left, and bottom right respectively

20.3.3 Time scales analysis of the diffusion impinging flame

Image capturing itself is an image processing process due to the limited exposure time (or gating time). It is not possible for a capturing process to resolve flame time scales which are smaller than the camera exposure time. Therefore image capturing is a time averaging process. The content of an image is an ensemble of various flame stages, within the chosen camera exposure time. Longer exposure times, i.e. low shutter speed, result in an averaged flame pattern and shorter exposure times give a more instantaneous look of a flame.

Figure 20.14 shows images of a ‘ring’ flame formed by a pure fuel central jet and a co-flow air jet impinging on a plate. The global fuel-to-air equivalence ratio is 2 and the ratio of the plate to nozzle separation distance (H) over the air nozzle diameter (D) is H/D = 3. A range of shutter speeds (1/30th second to 1/1600th second) was used. It can be seen that the averaged flame patterns, i.e. at low shutter speed, are very different from the more ‘instantaneous’ flame patterns captured at high shutter speed. It is obvious that an image at low shutter speed has ‘gathered’ more flames than at the high shutter speed.

Fig. 20.14 ‘Ring’ flame images captured by high-end digital camera of different shutter speeds

20.4 Conclusions

Flame visualization enhancement using digital image processing has been demonstrated. Traditional digital image processing techniques such as contrast enhancement and noise reduction are very valuable in improving the visualization quality. It was found that image subtraction between consecutive images are effective in identifying otherwise hidden flame structures. However the time separation between the two images has to be considered carefully. Contour tracking of the flame boundaries is useful in visualizing the dynamics and extent of flame movement. Further quantitative information extraction of the processed flame boundaries would give an indication on the global characteristic of each flame. The variable shutter speeds of modern digital camera are a valuable asset for the studies of flame time scales.


(1) Foat, T., Yap, K. P., and Zhang, Y., Combustion and Flame, Vol. 125, pp. 839–951, 2001.

(2) Zhang, Y., Rogg, B., and Bray, K. N. C., Combust. Sci. and Tech, Vol. 113–114, pp. 255–271, 1996.

(3) Jähne, B., Digital Image Processing – Concepts, Algorithms, and Scientific Applications, 4th Edition. Springer, 1997.

(4) Veynante, D., Duclos, J. M., and Piana, J., Twenty-Fifth Symposium (International) on Combustion, The Combustion Institute, Pittsburgh, pp. 1249–256, 1994.

(5) Deschamps, B. M., Smallwood, G. J., Prieur, J., Snelling, D. R., and Guider, Ö. L., Twenty-Sixth Symposium (International) on Combustion, The Combustion Institute, Pittsburgh, pp. 427–435, 1996.


Pope, S. B. Int. J., Engng. Sci. Vol. 26, pp. 445–469, 1988.

Zhang, Y. Experiments in Fluids, S282–S290, pp. 282–290, 2000.

W B Ng, K Y Cheung, and Y Zhang

Mechanical, Aerospace and Manufacturing Engineering Department, UMIST, Manchester, UK