You are on page 1of 8

# Local Binary Pattern

Pi19404
January 18, 2014

Contents

Contents
Local Binary Pattern 3

2|8

## Local Binary Pattern

     
In this article we will look at concept of Local Binary Pattern and computation of LBP image. 2D surface texture is characterize by spatial pattern and intensity/contrast. Spatial Pattern is affected by rotation,scale changes ,hence for a good texture description we require a rotation and scale invariant descriptor. Local binary pattern binarizes the local neighborhood of each pixel and builds a histogram on these binary neighborhood patterns. Let P be the number of neighborhood pixels and R the distance from the center pixel l and l be neighborhood pixel.
c p

A LBP number characterizes the local texture by assigning the binomial factor 2 for each sign sgn(l l )
P;R P P

LBPP;R

X
p

1
sgn(lp

=0

lc )2

   

lp

for p = 0 : : : P of radius R.
LBPP;R

## 1 are a set of equally spaced pixels on a circle

P

features has 2 possible values.For P=8 we have a binary feature vector of length 256.

Patterns are classified as uniform and non uniform. Uniform pattern have single contigious regions of 0 and 1 while non uniform patterns do not.For example 01100000 is a uniform pattern while 01010000 is a example of non uniform pattern

3|8

## Local Binary Pattern

we can see that there are 9 possible uniform pattern values 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 1 0 0 0 1 1 1 1 1 1 0 0 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 Now consider the effect of rotation on the feature vector. Rotating the image results in circular shift of values of feature vector. To encorporate rotational invariance we need to assign all possible rotation of a feature vector to a single LBP value For example all the below patterns 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 will be assigned to a single value. In the present article however we are not considering rotational invariant features Let us consider the implementation details of LBP If we only consider a 3x3 neighborhood we need to threshold the rectangular region about Another method would be to divide image into square blocks of size BxB.Instead of central pixel value we consider the mean value of pixels in the central block. Similariy instead of considering the single pixel value in the neighborhood we would consider the mean value of pixels in the block. All the pixels in the block are encoded with the same binary value 0 or 1.

  

     

4|8

## Local Binary Pattern

 

to compute mean value over rectangular regions of image,integral images are used. output of lbp images for block size 1,2 and 8 is showing in figure 1

## Figure 1: LBP Images

we can see that as block size increases,quantization effects can be seen and the information in the encoded image cannot be recognized.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

nclude "ImgFeatures/integralImage.h" ass LBPFeatures blic: LBPFeatures(){}; Mat image; vector<uchar> features; Mat mask; IntegralImage ix; //function to compute LBP image with block size 1 void compute(Mat image,Mat &dst) { uchar *ptr=image.data; image.copyTo(dst); uchar *optr=dst.data;

5|8

## Local Binary Pattern

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61

int width=image.cols; int height=image.rows; for(int i=1;i<height-1;i++) { for(int j=0;j<width-1;j++) { int center=(int)ptr[j+i*width]; unsigned char code=0; code|=((int)ptr[(j-1)+(i-1)*width] >=center)<<7 ; code|=((int)ptr[j+(i-1)*width] >=center)<<6 ; code|=((int)ptr[(j+1)+(i-1)*width] >=center)<<5 ; code|=((int)ptr[(j+1)+(i)*width] >=center)<<4 ; code|=((int)ptr[(j+1)+(i+1)*width] >=center)<<3 ; code|=((int)ptr[j+(i+1)*width] >=center)<<2 ; code|=((int)ptr[j-1+(i+1)*width] >=center)<<1 ; code|=((int)ptr[j-1+(i)*width] >=center)<<0 ; optr[j+i*width]=code; } } } //computing integral image for block void computeBlock(Mat image,Mat & dst,int block=2) { /computing integral image ix.compute(image); image.copyTo(dst); dst.setTo(cv::Scalar::all(0)); int width=image.cols; int height=image.rows; for(int i=block;i<height-block;i=i+block) { for(int j=block;j<width-block;j=j+block) { int x=i; int y=j; Rect r=Rect(j,i,block,block); //computing mean value of central block int meanv=ix.calcMean(r); int code=0; for(int k=0;k<8;k++) {

6|8

## Local Binary Pattern

62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105

//offsets of neighbouring values switch(k) { case 0: y=i-block; x=j-block; break; case 1: y=i; x=j-block; break; case 2: y=i+block; x=j-block; break; case 3: y=i+block; x=j; break; case 4: y=i+block; x=j+block; break; case 5: y=i; x=j+block; break; case 6: y=i-block; x=j+block; break; case 7: y=i-block; x=j; break; default: break; } Rect r1=Rect(x,y,block,block); //computing mean value of neighbouring block int val=(int)ix.calcMean(r1); code|=(meanv >= val)<<(7-k); }

7|8

## Local Binary Pattern

106 107 108 109 110 111 112 113 114 115 116

} }

Mat roi=dst(r); //setting value of all pixel in output //image to the encoded value for visualization roi.setTo(cv::Scalar::all(code));

the code for the same can be found in the git repo for OpenVisionLibrary https://github.com/pi19404/OpenVision/ in following files ImgFeatures/lbpfeatures.hpp and lbpFeatures.cpp files

8|8