'radius' is the distance from the center pixel to the neighbours, 'neighbours' is their count, like here:
with lbph, the face region gets divided into seperate patches, like this:
grid_x and grid_y are the number of patches in x and y direction (8 in this case, like the default with opencv)
for each patch, a separate (2^neighbours bins)histogram is calculated,
and later concatenated to a
grid_x * grid_y * (2^neighbours) element feature vector.
(you can already see, that increasing the neighbours count will be very expensive)
[images taken from : Face Recognition with Local Binary Patterns,
Timo Ahonen, Abdenour Hadid, and Matti Pietikäinen - you might want to read this..]