Large pose face alignment is a prerequisite for many important visual tasks such as face recognition and 3D face reconstruction. However, most of the existing alignment methods use two-dimensional boundary markers to align, and the number of boundary markers used is limited, which greatly affects the accuracy of large pose face alignment. Therefore, an improved large pose face alignment method is proposed. Firstly, 3D deformable model is utilized to represent 2D face image. And the problem of face alignment with arbitrary pose is modeled as a fitting problem based on Three Dimensional Deformation Model(3DMM). And then a cascade regression method based on Convolutional Neural Network(CNN) is adopted to learn the mapping relationship between two-dimensional face image and its representation. Finally, two new pose invariant local features are proposed as the input layer of CNN learning, and CNN is applied for large pose face alignment through training. Simulation results on two classic face image data sets show that the proposed method is better than the latest face alignment method.