Robust Face Super-Resolution via Patch Network of Global Context Prior

Abstract

The face image captured in real-world scene is the vital object in visual tasks due to its importance in person identity confirmation. However, the face structures is inherently missing due to the degradations in the imaging process, leading to the severe interference in the face enhancement and subsequent face recognition procedure. To remove the degradations and recover a recognizable face, we propose a novel global facial context prior as the guidance to face super-resolution problem. The global facial context prior expands the local structure prior, which is represented as the local patches, into global-face range by modeling the whole facial patch positions into network scheme empirically according to position distributions. The local structure in neighbored positions provides better guidance for reconstruction than the local structure itself, especially when the size of local distortion is larger than the local patch size in input image, leading to the fact that the local input image patch unable to give informative details. Therefore, the global facial context prior improves the robustness of algorithm in dealing with large-range uneven image distortion, e.g. the occlusion. The quantitative and qualitative evaluation on public database demonstrates the superiority of our algorithm.