This paper presents an improved local binary pattern (LBP) descriptor, known as the local mean pattern (LMP), to classify static gestures. We select original gesture images, nonlinear illumination images,Gaussian blurred images, and images contaminated by salt and pepper noise, calculate their LMP, LBP and
local angular phase (LAP) descriptor. The gentle_Adaboost classification algorithm is used for training and verifying these gesture features. The descriptor makes full use of correlation and difference of pixel gray values in certain regions. It is a good description for the characteristics of different gesture images. Experiment results show that LMP descriptor outperforms LBP and LAP. Classification accuracy of LMP descriptor reaches 95%.  The descriptor is robust to nonlinear illumination and Gaussian blur.