This paper describes an automatic face recognition algorithm for security entrances. There are two major steps in this procedure to make the automatic recognition possible: (1) We combined the two-phase face detection method and back propagation neural networks to detect human faces when people are walking in the region of entrances. The combination allows the strength of both methods activated to accommodate the size and head-orientation variations and to eliminate the false detection. (2) Novel face recognition: we extract the facial feature measurements to form the multi-variable normal distribution for each person. These multi-variable normal distributions separate the decision space well and the probability for good index for face recognition. This face recognition algorithm is very efficient on computing time and taking little storage space.