Research on Face Restoring in Surveillance Videos

Abstract

Since cases being analyzed and detected by the police using surveillance videos, insufficient resolution and degradation always lead to the suspect’s face being unrecognized. For such cases, from the perspective of learning-based face superresolution method, this paper proposed a novel face restored solution based on face super-resolution and image mosaic. Firstly, by applying global similarity selection over the entire database, the scope of the used samples can be narrowed and boost the speed. Then the local representation between neighbors can make the method more robust against noise under severe conditions. Finally the reconstructed face could be inserted into original background by the improved Poisson Editing algorithm. This scheme achieve the effect of face restoring on a single frame in surveillance videos. Experiments under common database and real surveillance video scenarios demonstrate our method has better performance, especially for noise input images.