Occlusion detection for ICAO compliant facial photographs

Facial image analysis is an important computer vision topic as a first step for biometric applications like face recognition/verification. The ICAO specification defines criteria to assess suitability of facial images for later use in such tasks. This standard prohibits photographs showing occlusions, thus there is the need to detect occluded images automatically. In this work we present a novel algorithm for occlusion detection and evaluate its performance on several databases. First, we use the publicly available AR faces database which contains many occluded face image samples. We show a straight-forward algorithm based on color space techniques which gives a very high performance on this database. We conclude that the AR faces database is too simple to evaluate occlusions and propose our own, more complex database, which includes, e.g., hands or arbitrary objects covering the face. Finally we extend our first algorithm by an Active Shape Model in combination with a PCA reconstruction verification. We show how our novel occlusion detection algorithm outperforms the simple approach on our more complex database.