Abstract: Deep learning is currently an extremely active research area in machine learning and pattern recognition society.Deep learning and Big Data are two hottest trends in rapidly growing digital world.Today，big data for the development of various industry has brought the huge opportunity and potential.On the other hand it has also brought unprecedented challenges.While big data has been defined in different ways，it is referred to the exponential growth and wide availability of digital data.It is difficult or even impossible to manage and analyze them using conventional software tools and technologies.It has gained huge successes in a broad area of applications such as speech recognition，computer vision and natural language processing.In the field of computer vision optical character recognition (OCR) concept was put forward in early 1920s.It is a representative in the field of pattern recognition research important topic.Recently，RBM(Restricted Boltzmann Machines) becomes increasingly popular because of its fast learning algorithm，Contrastive Divergence.For the theoretic perspective，the success of RBM has greatly encouraged the research about the stochastic approximation theory，energy­based models and unnormalized statistical models.For the application perspective，RBM has been successfully applied in various machine learning domains，such as classification，regression，dimension reduction，high­dimensional time series modeling，sparse over­complete representations，image transformations and collaborative filtering.In this article，we proposed a new compositional DBM­ELM network model based on deep Restricted Boltzmann machines which realize data nonlinear dimension reduction and Extreme Learning Machines algorithm.This model can simplify the high­dimensional data into low­dimensional space better and then get a better classification effect to represent complicated functions.Finally the experiments on face and handwritten digit recognition prove the superiority.