Extending a local matching face recognition approach to low-resolution video

Identifying persons in surveillance videos by automatic face recognition is a difficult task, caused by poor image resolution among other things. For high-resolution face data, local matching approaches have proven to achieve better results than holistic ones. However, for low-resolution videos, the holistic approaches are the most widely used solution because the scale can be changed easily. Whereas, the local matching approaches are not commonly used as the decreasing size of the local regions raises difficulties. With local binary patterns (LBP) as feature for local matching, we address this problem by suggesting several modifications. By using different scales and temporal fusion, we can avoid sparse LBP-histograms in the small local regions even for low resolutions. Following this concept, the application range of the local matching approach is extended down to faces with a size of 8× 8 pixels. Reaching this scale enables face recognition in low-resolution surveill ance videos.