International Journal of Current Engineering and Technology

E-ISSN 2277 – 4106, P-ISSN 2347 - 5161

®
©2014 INPRESSCO , All Rights Reserved
Available at http://inpressco.com/category/ijcet

Research Article
Smoke Detection using Local Binary Pattern
D.A AgrawalȦ and Prachi MishraȦ*
Ȧ
Shri Sant Gadge Baba Collage of Engg. & Technology Bhusawal, India

Accepted 30 Nov 2014, Available online 01 Dec 2014, Vol.4, No.6 (Dec 2014)

Abstract

To realize quick and robust fire detection with image information of real scenes, smoke is a key feature information in
detection methods. Since smoke does not keep stationary shape, it is difficult apply ordinal image processing techniques
such as the edge or contour detection directly. Image information of smoke is also affected from its environmental
conditions such as illumination changes and background objects. In this study, we adopt Local Binary Patterns (LBP),
which is defined as a simple texture operator computed using the center pixel value and its neighborhood pixel values.
From its definition, LBP is a robust image descriptor against the illumination change. The adaptive detection for real–
scene situations is realized by AdaBoost. Results using with real scene data show that the presented method can provide
accurate results against the various conditions of real world situation.

Keywords: Smoke detection, Binary Pattern etc.

1. Introduction serious problem for open area image processing as

whether conditions, for examples, clouds, lighting sources,
1
For fire alert system, the most important problem is how wind, etc., are always changing dynamically.
quick the existence of fire is detected. The fire detection in LBP is defined as a simple texture operator computed
open areas, image information is mainly used due to the using the center pixel value and its neighborhood pixel
fact that other sensing devices such as a gas senser become values. As it is considering the result as a binary number,
high– cost and cannot set them in some places. Some LBP is robust against the illumination change. we prepare
methods of detecting a fire or flame directly (Hidenori a training set, we need to consider that the image
Maruta et al, 2013; H. Maruta et al, 2011; H. Maruta et al, information of smoke is strongly affected from its
2010) are proposed, however, they have some difficulties background. That is, we have difficulties to gather smoke
to catch fire with cameras depending on the relation of images and non–smoke images in a wide range of
places between a camera position and a fire generating variations. We consider smoke detection as a two–class
place. problem in the context of learning problem.
For such purpose, smoke is a key feature since we can
observe it even if fire is not in the field of view of the 2. Smoke Detection Using LBP
camera. There are some computer vision based methods
for smoke detection which use the edge or color A. Preprocessing
information based technique (B.U.Toreyin, Y.
Dedeoglu13th European Signal Processing Conference, In the presented method, we detect moving objects in an
Antalya, Turkey, 2005.) They have difficulties in treating im- age sequence as candidates of smoke regions in a
characteristic properties of smoke and needing high-cost preprocesing. As a growth–speed of smoke is considered,
computations to detect smoke whole of wide-view images we obtain 1 frame–per–second rate image sequences {(𝑡)∣
or image sequences. There are other approaches which 𝑡 =0 ,1,2,⋅⋅⋅} from original image sequences obtained from
combine texture features and its time–series properties (H. a single camera. We use an image subtraction technique to
Maruta, A. Nakamura and F. Kurokawa et al 2010 IEEE ). extract regions of moving objects, which are candidates of
It is shown that they can provide accurate results of smoke smoke. The subtracted image frame is written as
detection in open areas. However, they need relatively (𝑡)=𝑓(𝑡)−𝑓(𝑡−1). This(𝑡) is used in a following main
long time to obtain the result as they use the time processing.
accumulation technique, which becomes the problem in
some real world situations. Additionally, these methods B. Local Binary Patterns
have a common problem, which is that the detection result
is affected from an illumination change. This becomes a Local binary patterns of LBP deals with the neighbor
pixels in an image by using thresholding the neighborhood
*Corresponding author: Prachi Mishra of each pixel and consider the result as a binary number,
4052 | International Journal of Current Engineering and Technology, Vol.4, No.6 (Dec 2014)
D.A Agrawal et al Smoke detection using LBP

and thus, is used for segmentation and clustering in an neighborhood. The basic LBP using 8 pixels in a 3×3 pixel
image. The Algorithm is as follows block, this generic formulation of the operator puts no
 The LBP feature vector, in its simplest form, is limitations to the size of the neighborhood or to the
created in the following manner: number of sampling points.
 Divide the examined window into cells (e.g. 16x16 After the preprocessing, we compute local binary
pixels for each cell). patterns (LBP) from subtracted images (𝑡). LBP is a
 For each pixel in a cell, compare the pixel to each of simple tex- ture operator computed using the center pixel
its 8 neighbors (on its left-top, left-middle, left- value and its neighborhood pixel value (T. Ojala, M.
bottom, right-top, etc.). Follow the pixels along a Pietika¨ainem and D. Harwood: et a1996.). Previous
circle, i.e. clockwise or researches showed that LBP operator works well in both
Where the center pixel's value is greater than the reducing computational complexity and reasonable feature
neighbor's value, write 1. Otherwise, write 0. This gives an extraction of images com- pared to conventional texture
8-digit binary number (which is usually converted to features (T. Ojala, M. Pietik¨ainen and T. M¨aenp¨a¨a:
decimal for convenience). IEEE Trans. et al 2002),. Especially, it is suggested that
 Compute the histogram, over the cell, of the LBP is effective to illumination.
frequency of each number occurring (i.e., each
combination of which pixels are smaller and which
are greater than the center).
Optionally normalize the histogram

Fig.4 The Original LBP

Fig.1 LBP based texture identification Problem

The LBP was proposed as a two-level version of the

texture unit to describe the local textural patterns The
original version of the local binary pattern operator works
in a 3×3 pixel block of an image. The pixels in this block
are thresholded by its center pixel value, multiplied by
powers of two and then summed to obtain a label for the
center pixel.

Fig. 5 Computation of LBP

The smoke detection in open areas, the lighting condition

always changes. Thus, we consider that the technique
using LBP is effective in smoke detection in open areas
which has influence of environmental conditions such as
Fig.2 Example of an input image. The corresponding LBP weather, day or night time, and so forth. In this study, we
image and Histogram adopt the original LBP value (T. Ojala, M. Pietika¨ainem
and D. Harwood: et al 1996.). The original LBP is simply
computed using the center pixel value and its
neighborhood pixel value .The LBP operator works in a
3×3 neighborhood, using the center pixel value as a
threshold. The neighborhood pixels value of the center
pixel are multiplied by the binomial weights given to the
corresponding pixels and obtained values are summed for
LBP value (LBP = 1⋅20 + 0⋅21 + 0⋅22 + 0⋅23 + 1⋅24 +
Fig.3 The Circular (8,1), (16,2) And (8,2) Neighborhoods. 0⋅25 + 1⋅26 + 1⋅27 = 209) of this texture unit. Using LBP
The Pixel Values Are Bilinearly Interpolated Whenever values from an image block divided into 24 × 24 pixels,
The Sampling Point is not in the Center of a Pixel we create a LBP histogram. The above process to compute
the LBP histogram is shown in Fig. 2. We compute the
As the neighborhood consists of 8 pixels, a total of 28 LBP histogram of each image block and treat it as a 256-
=256 different labels can be obtained depending on the dimensional feature vector of the image block. Therefore,
relative gray values of the center and the pixels in the the detection result is block–based one.
4053 | International Journal of Current Engineering and Technology, Vol.4, No.6 (Dec 2014)
D.A Agrawal et al Smoke detection using LBP

The local binary pattern operator is an image operator ξm(H(i)) = {

which transforms an image into an array or image of
integer labels describing small-scale appearance of the +1 (δs-δN >Ѳm)
image. These labels or their statistics, most commonly the -1 (otherwise) (3)
histogram, are then used for further image analysis. The
most widely used versions of the operator are designed for where 𝜃𝑚 is the threshold value adjusted to every iteration
monochrome still images but it has been extended also for step 𝑚 of AdaBoost so that an error rate serves minimum
color (multi channel) images as well as videos and value when the sample data is identified. When the input
volumetric data. The actual LBP operator in spatial LBP histogram ( ) is similar to smoke, it is classified to
domain ,while deals with spatiotemporal LBP . smoke (+1), else it is classified to non–smoke (−1), and
When we use the LBP histograms as a feature vector, then we attach a label (smoke or non–smoke) to the image
the components of input vector consist of each histograms block. The weak classifier is defined by this labeling. That
bin. We define the similarity of the LBP histogram which is, the weak classifier is defined by the similarity measure
measure the distance from smoke–class and non–smoke with the Bhattacharyya coefficient (Eq. (1) and (2))and the
class with Bhattacharyya coefficient (K. Fukunaga: et al, classification result is obtained from Eq. (3)
1990.)
We prepare the averaged LBP histogram 𝐻𝑆 and 𝐻𝑁 3. Smoke Detection Method Using LBP
from manually selected smoke and non–smoke image
blocks. The Bhattacharyya coefficient represents the  Image Pre-Processing for Motion Detection (Smoke)
similarity of two histograms in 0 to 1. In this study, using direct Subtraction
Bhattacharyya coefficients of input LBP histogram and  Local Binary Pattern based Map Generation using
averaged histograms of both Histogram
Ada-Boost algorithm based Artificially Intelligent
Learning implementation over thousands of images of the
same place with smoke.

Proposed Algorithm

Fig.7 Flow of Algorithm

4. Image Segmentation using 2D Osto Method

Image segmentation is one of the basic techniques of

Fig.6 Calculation of Bhattacharyya Coefficients As The image processing and computer vision. It is a key step for
Distance From Smoke Class and Non–Smoke Class. image analysis, comprehension and description. Among
all the segmentation techniques, thresholding
classes are written as Eq. (1) and (2) segmentation method is the most popular algorithm and is
widely used in the image segmentation field. The basic
δs(Hs ,H(i)) = ∑ √𝐻 𝐻 (1) idea of automatic thresholding is to automatically select an
optimal or several optimal gray-level threshold values for
δN(HN ,H(i)) =∑ √𝐻𝑁𝐻 (2) separating objects of interest in an image from the
background based on their gray-level distribution.
where 𝐻𝑆 =( 𝐻𝑆 1 ,⋅⋅⋅,𝐻𝑆256) and 𝐻𝑁 =( 𝐻𝑁 1 ,⋅⋅⋅,𝐻𝑁 256) are Automatic thresholding techniques can be roughly
average histograms of smoke and non–smoke, 𝐻( ) = (𝐻( ) categorized as global thresholding and local thresholding.
() Otsu thresholding technique is one of the global
1 ,⋅⋅⋅,𝐻 256) is an -th histogram of image block obtained
from an input image. This calculation process is shown in thresholding method and has been cited as an effective
Fig. 3. using Eq. (1) and (2). technique . One of them is its sensitivity to the object size
We define weak classifiers on the basis of the Say in brief, if the object proportion is much less than
difference of similarity of the Bhattacharyya coefficient background, the pixels in background will be wrongly
(𝐻𝑆,( )) and 𝛿𝑁(𝐻𝑁,𝐻( )) as shown in Eq. (3) classified as object; on the contrary, if the object
4054 | International Journal of Current Engineering and Technology, Vol.4, No.6 (Dec 2014)
D.A Agrawal et al Smoke detection using LBP

proportion is much more than background, the pixels in 6. Avantages

object will be wrongly classified as background. As for
our renal biopsy samples, the object size is much less than Most of the algorithms do not take care of changing values
background, the wrong classification of pixels by of brightness due to change in weather conditions, and
traditional Otsu method will lead to the failure thus are not so practical for open area’s, like forests, and
segmentation hills etc. (Forest fires being a very big issue in United
States).
No algorithm has tried to estimate the amount of
5. Threshold Selection Method from Gray Level smoke getting generated, so as to make an early warning
Histogram system. If multiple camera’s are synchronized together,
then the camera’s away from the source will register more
A nonparametric and unsupervised method fautomatic smoke, but are not useful for ringing an alarm, as they do
threshold selection for picture segmentation is presented. not have any information about the source of the fire
An optimal threshold is selected by the discriminator
criterion, namely, so as to maximize the separability of the Conclusion
resultant classes in gray levels. The procedure is very
simple, utilizingonly the zeroth-and the first-order Smoke detection method based on local binary patterns. In
cumulative moments of the gray-level histogram. It is smoke detection, LBP is not only a simple operator but
straightforward to extend the method to multithreshold also effective to illumination variations, which can
problems. Several experimental results are also presented contribute to obtain accurate results. We also combine
to support the validity of the method. with AdaBoost, which is one of the machine learning
It is important in picture processing to select an techniques, to improve the accuracy of detection results.
adequatethres hold of gray level for extracting objects As the preprocessing, we extracted the moving objects as
from their background. A variety of techniques have been candidate smoke regions in images. From the subtracted
proposed in this regard. In an ideal case, the histogram has images, we computed LBP values and LBP histograms
a deep and sharp valley between two peaks representing from the image blocks. Using LBP histograms as the input
objects and background, respectively, so that the threshold vector of AdaBoost, we detect whether the image blocks
can be chosen at the bottom of this valley . However, for are smoke or non–smoke.
most real pictures, it is often difficult to detect the valley
bottom precisely ,especially in such cases as when the Acknowledgment
valley is flat and broad, imbued with noise, or when the
two peaks are extremely unequal in height, often I cannot complete my work without support of my internal
producing no traceable valley. There have been some guide.
techniques proposed in order to overcome these
difficulties. They are, for example, the valley sharpening References
technique, which restricts the histogram to the pixels with
large absolute values of derivative (Laplacian or gradient), Hidenori Maruta, Yusuke Iida, Fujio Kurokawa (2013),
and the difference histogram methode ,which selects the Smoke Detection Method Using Local Binary Patterns and
threshold at the gray level with hexadecimal amount of AdaBoost, IEEE paper.
difference. These utilize information concerning H. Maruta, Y. Kato and F. Kurokawa (2011), Smoke
neighboring pixels (or edges) in the original picture to detection using time series property of texture features,
modify the histogram so as to make it useful for IEEE Trans. Electrical and Electronic Engineering Volume
thresholding. Another class of methods deals directly with 6, Issue 4, pp. 390–39
the gray-level histogram by parametric techniques. For H. Maruta, A. Nakamura and F. Kurokawa (2010), A new
example, the histogram is approximated in the least square approach for smoke detection with texture analysis and
sense by a sum of Gaussian distributions, and statistical support vector machine, Proceedings of 2010 IEEE
decision procedures are applied . However, such a method International Symposium on Industrial Electronics, pp.
requires considerably tedious and sometimes unstable 1550 – 1555.
calculations. Moreover, in many cases, the Gaussian B. U. T¨oreyin, Y. Dedeoglu and A. E. Enis (2005) Wavelet
distribution sturn outtobeameager approximation of the based real–time smoke detection in video, 13th European
real modes. In any event, no goodness of threshold has Signal Processing Conference, Antalya, Turkey.
G. Zhao and M. Pietika ¨inen (2007.), Dynamic Texture
been evaluated in most of the methods so far proposed.
Recognition Using Volume Local Binary Patterns, Proc.
This would imply that it could be the right way of deriving
Workshop Dynamical Vision, pp. 165-177
an optimal thresholding method to establish an appropriate H. Maruta, A. Nakamura and F. Kurokawa (2010)A new
criterion forevaluatingthe goodness of threshold from a approach for smoke detection with texture analysis and
more general standpoint. It is not only important as a support vector machine, Proceedings of 2010 IEEE
standard technique in picture processing, but also essential International Symposium on Industrial Electronics, pp.
for unsupervised decision problems in pattern recognition. 1550 - 1555.
A new method is proposed from the viewpoint of K. Fujita and S.K. Nayar (2003), Recognition of Dynamic
discriminant analysis; it directly approaches the feasibility Textures Using Impulse Responses of State Variables,
of evaluating the goodness of threshold and automatically Proc. Third Int’l Workshop Texture Analysis and
selecting an optimal threshold. Synthesis. pp. 31-36
4055 | International Journal of Current Engineering and Technology, Vol.4, No.6 (Dec 2014)
D.A Agrawal et al Smoke detection using LBP

T. Ojala, M. Pietika¨ainem and D. Harwood (1996) A M. Pietika¨ainen, G. Zhao, A. Hadid and T. Ahonen (2011),
comparative study of texture measures with classification Computer vision using local binary patterns, Springer,
based on feature distributions, Pattern Recognition, Vol. London
29, pp. 51-59. B. U. T¨oreyin, Y. Dedeoglu and A. E. Enis (2005)Wavelet
T. Ojala, M. Pietik¨ainen and T. M¨aenp¨a¨a (2002) based real–time smoke detection in video, 13th European
Multiresolution gray-scale and rotation invariant texture Signal Processing Conference, Antalya, Turkey.
classification with local binary patterns, IEEE Trans. Rahul Sasidharan, Menaka D (April 3-5, 2013) Dynamic
Pattern Analysis and Machine Intelligence, Vol. 24, pp. Texture Segmentation of Video Using Texture Descriptors
971–987 and Optical Flow of pixels for Automating Monitoring in
K. Fukunaga (1990) Introduction to statistical pattern Different Environments, International conference on
recognition, Morgan Kaufmann, San Francisco. Communication and Signal Processing.
G. Doretto, A. Chiuso, S. Soatto, and Y.N. Wu (2003), G. Zhao and M. Pietika ¨inen (2007.), Dynamic Texture
Dynamic Textures, Int’l J. Computer Vision, vol. 51, no. 2, Recognition Using Volume Local Binary Patterns, Proc.
pp. 91-109. Workshop Dynamical Vision, pp. 165-177
H. Schneiderman and T. Kanade (2004) Object Detection
Using the Statistics of Parts, International Journal of
Computer Vision, Vol. 56, pp. 151–177

4056 | International Journal of Current Engineering and Technology, Vol.4, No.6 (Dec 2014)