Local Binary Patterns

LBP turns every pixel into an 8 bit texture code from its neighbors, so faces and surfaces match by pattern counts rather than raw brightness.

LBP thresholds eight neighbors into bits and pools the codes into a texture histogram.

Why Does This Exist?

Texture tells faces, wood, and fabric apart when color and exact brightness do not. Raw pixels shift with every lamp change, but the order of brighter versus darker neighbors stays stable. LBP captures that order cheaply.

It codes each pixel from its ring, then pools codes into histograms over cells. Those histograms feed classical classifiers for faces and materials. For edge direction histograms instead, see HOG.

Think of It Like This

A neighborhood height census

Stand on one paving stone and ask which of the eight around you stands taller. Write yes or no eight times. That 8 bit note describes your spot without naming any height.

Do this for every stone, then count note types per block. Two brick paths give similar counts even when one sits in shade. The analogy stops at scale: fixed rings miss larger weaves unless you grow the radius.

How It Actually Works

For center $g_c$ and $P = 8$ neighbors $g_p$ on radius $R = 1$ , set bit $p$ when $g_p \ge g_c$ :

$\mathrm{LBP}_{8,1} = \sum_{p=0}^{7} s(g_p - g_c) 2^p, \quad s(z) = 1 \text{ if } z \ge 0$

This gives values $0$ to $255$ . Uniform patterns with at most two $0$ to $1$ transitions keep $58$ codes plus one bin for the rest, so $59$ bins per cell. Histograms over cells form the face or texture vector.

Worked numbers

Center $100$ with neighbors $[110, 95, 102, 99, 120, 88, 101, 105]$ gives bits $[1, 0, 1, 0, 1, 0, 1, 1]$ , which sums to $1 + 4 + 16 + 64 + 128 = 213$ . That code has six transitions, so it falls in the non-uniform bin. A flat cheek patch of $[101, 100, 102, 99, 101, 100, 101, 100]$ gives mostly ones, code $255$ , a uniform spot. Face grids often use $7 \times 7$ cells times $59$ bins for $2891$ values.

Watch Out For

Noise flips single bit codes

One gray level of noise can flip a neighbor bit and move a pixel to a distant code. Smooth lightly or use ternary variants when sensor noise is high, or histograms jitter frame to frame.

Fixed radius on multiscale texture

Radius $1$ sees yarn but not weave. Add radii $2$ and $3$ or a pyramid when texture spans sizes, or large patterns alias into uniform bins.

The Quick Version

LBP codes each pixel from 8 neighbor comparisons.
Uniform codes compress $256$ values to $59$ bins.
Cell histograms form a lighting tolerant texture vector.
Strong for faces and materials, weak across large scale shifts.
Pair multiscale radii for coarse weaves.