An attempt to encode images in a way which preserves texture information and minimises lighting and other pesky variations.

Image data:

| 5 | 4 | 3 |
| 4 | 3 | 1 |
| 2 | 0 | 3 |

Threshold with respect to the centre pixel:

| 1 | 1 | 1 |
| 1 |   | 0 |
| 0 | 0 | 1 |

Binary encode with:

| 1 | 2 | 4 |
| 8 |   | 16 |
| 32 | 64 | 128 |

To produce:

| 1 | 2 | 4 |
| 8 |   | 0 |
| 0 | 0 | 128 |

= 143

This number represents the relationship of the centre pixel with it's neighbours in a way which ignores overall changes in contrast or level. 

Obviously it's not too instructive to look at these images, but obviously we are going to try:

{{:lbp.png|}} {{:lbp2.png|}}

Further stuff:

  * For face detection/recognition: [[http://www.jdl.ac.cn/doc/2005/Local%20Gabor%20Binary%20Pattern%20Histogram%20Sequence%20(LGBPHS)%20A%20Novel%20Non-Statistical%20Model%20for%20Face%20Representation%20and%20Recognition.pdf]]