Face hallucination based on morphological component analysis

In this paper, we formulate the face hallucination as an image decomposition problem, and propose a Morphological Component Analysis (MCA) based method for hallucinating a single face image. A novel three-step framework is presented for the proposed method. Firstly, a low-resolution input image is up-sampled via an interpolation. Then, the interpolated image is decomposed into a global high-resolution image and an unsharp mask by using MCA. Finally, a residue compensation is performed on the global face to enhance its visual quality. In our proposal, the MCA plays a vital role as MCA can properly decompose a signal into several semantic sub-signals in accordance with specific dictionaries. By virtue of the multi-channel decomposition capability of MCA, the proposed method can be also extended to simultaneous implementation of face hallucination and expression normalization. Experimental results demonstrate the effectiveness of our method for the images from both lab environment and realistic scenarios. We also study the contribution of face hallucination to face recognition in the case that probe images and gallery images are under different resolutions. The main conclusion is that the contribution is significant when using local facial features (e.g., LBP), but unobvious when using holistic facial features (e.g., Eigenfaces).

► The face hallucination is formulated as an image decomposition problem.
► A Morphological Component Analysis is used to solve the decomposition problem.
► Face hallucination and expression normalization can be solved simultaneously.
► The contribution of face hallucination to face recognition is studied.