Hallucinating faces in curvelet

Abstract

In this paper, we aim to enhance the resolution of a single face image. We introduce a method which utilizes the specific features of Curvelet to select training samples and estimate local face features. Based on different characteristics of Curvelet coarse and fine coefficients, we firstly set up training sets for global face enhancement and local features estimation separately. Secondly, global faces are derived by employing sparse representation technique. Thirdly, we transfer the low-resolution local features into Curvelet frequency domain and infer the relationship of Curvelet coefficients between testing images and training images. Through the learning process, the Curvelet coefficients of the high-resolution local features can be derived. Finally, the high-resolution local features are generated through Inverse Discrete Curvelet Transformation, which are then combined with the global face to produce the final hallucinated face. Experiment demonstrates that our approach outperforms other approaches.