Roughly, the main fundamental steps are the following:

- Get the face crop (maybe using a face detector) and align the face images (similarity or whatever technique you prefer) in order for the LBP you extract to be meaningful.

- Bring all the images to the same size or the resulting feature vectors will have different lengths and will not be matchable

- Decide the dimension of the cell size where the LBP will be calculated.

- Run an LBP extraction algorithm on the aligned face images. Concatenate all the LBPs extracted from the image cells to form a unique face feature vector

- Now that you have a feature vector for each face image, you can decide what to do. Usually at least a dimensionality reduction step is always performed (usual descriptors range from 10K to 100K dimensions). Than you can apply the learning or matching algorithm you prefer.

i hope this link useful to you 

http://docs.opencv.org/2.4/modules/contrib/doc/facerec/facerec_tutorial.html

you can refer my paper for more details about implementation of single scale, multiscale, block level etc. LBP features .

https://www.researchgate.net/publication/303408899_Color_Local_Phase_Quantization_CLPQ-_A_New_Face_Representation_Approach_using_Color_Texture_Cues

Conference Paper Color Local Phase Quantization (CLPQ)- A New Face Representa...

I think there is a common abuse of the term LBP for face recognition. What you should do is compute LBP histograms or histogram sequences. A good introduction into that is given by Ahonen et al. A free version can be found here:

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.214.6851