Feature extraction and pattern classification are two key problems in face recognition. In order to solve the high-dimensional and Small Sample Size (SSS) problem of face recognition, start with the feature extraction of human face and dimensionality reduction algorithms, a quadratic feature extraction and dimensionality reduction algorithm model was put forward based on Restricted Boltzmann Machine (RBM). At first, the image was evenly divided into a number of local image blocks and quantified, then the image was processed by Gabor wavelet transformation. The Gabor facial features were encoded by RBM to learn more intrinsic characteristics of data, so as to achieve the purpose of dimensionality reduction of high-dimensional facial features. On the basis of that, a multimodal face recognition algorithm based on Deep Belief Network (DBN) was proposed. The recognition results on ORL, UMIST and FERET face databases with different sample sizes and different resolution images show that, compared with the linear dimension reduction method and shallow network method, the proposed method achieves better learning efficiency and good recognition result.