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 $l_c$ and $l_p$ be neighborhood pixel.
- A $LBP_{P,R}$ number characterizes the local texture by assigning the binomial factor $2^P$ for each sign $sgn(l_p-l_c)$ \[ LBP_{P,R} = \sum_{p=0}^{P-1} sgn(l_p - l_c) 2^p \]
- $l_p$ for $p={0\ldots P-1}$ are a set of equally spaced pixels on a circle of radius $R$.
- $LBP_{P,R}$ features has $2^P$ 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
- 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
\\ will be assigned to a single value.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 - 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.
- 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 f1
- we can see that as block size increases,quantization effects can be seen and the information in the encoded image cannot be recognized.
- 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