A Survey of Human Face Detection
Liang Lü · Chinese Journal of Computers · 2002
This paper presents a survey on the state of the art of face detection research based on systematic analysis of related papers. Firstly face detection problem is divided into several classes according to the type of input images, background complexity, pose variance, application domain etc., and then face pattern is analyzed based on various features and their possible fusion method for the purpose of face detection. The literature is reviewed in two parts: feature extraction and feature fusion for face detection. Feature extraction includes skin color segmentation and various gray level features such as the outline of face, gray level distribution, organic feature, symmetry, template etc. Feature fusion methods include knowledge based heuristic face verification, statistical learning approaches (Eigenface, Clustering, ANN, SVM, HMM, EM probabilistic model). Performance comparison of some well known methods is given on MIT+CMU test set. In conclusion, statistical learning methods are superior to those knowledge based methods, and in all those learning based methods, the key problem is the training complexity, even by bootstrap method it remains a great challenge due to the diversity of non face samples compared with face samples. We suggest a subspace method for downsizing the training space by designing a filter (such as template matching filter) that excludes most of non face candidates and then training in the downsized subspace. It is pointed out that statistical learning methods depend on the accordance of sample patterns (syntactic information), which cannot take into considerations of much important semantic information. This differs much from human beings in face cognition. There is a limit for statistical only approaches and the help of knowledge based methods is needed.