Local binary pattern based face recognition with automatically detected fiducial points

摘要: This paper deals with automatic face recognition in the context of a real application for person identification developed Czech News Agency (y CTK) . We focus on popular Local Binary Patterns (LPBs) that are frequently used this field high accuracy. One drawback current LBP based methods is positions and number fiducial points fixed. These thus do not reflect properties particular image whereas we believe it beneficial to identify most representative ones. The main contribution consists proposing comparing several LBP-based approaches detect such fully automatically. use set Gabor filters task. extrema filter responses detected then as feature points. further reduced by clustering algorithm. Our also differ from other ones matching procedure. proposed evaluated three standard corpora: ORL, FERET, AR database our y CTK dataset containing uncontrolled images. Recognition results clearly show quality outperform significantly baseline approach all corpora. benefits particularly evident case non-controlled images corpus) where accuracy increased more than 20% absolute value.