An image based real-like virtual fitting system focusing on head accessories such as glasses and hat is proposed in this paper. The key techniques are 3D registration and virtual-real synthesis. Firstly, a facial landmark detection and pose estimation based algorithm for capturing registration data is presented. Then the approach using color blending and depth buffering to solve occlusion and model material problem is discussed. The registration algorithm is evaluated on the AFLW(Annotated Facial Landmarks in the Wild) dataset and testified to be precise enough for virtual fitting. Finally, a series of virtual fitting results are showed. The experimental results, which are obtained on the condition of pose variation and part occlusion, indicate that the proposed virtual fitting technique is fast, accurate and real-like.