Scale, Rotation and Translation Invarient Fast Face Detection System

H. Ikeda, N. Kato, H. Kashimura, and M. Shimizu (Japan)


Face detection, scale-rotation-translation invariant, face alignment detectors.


In this paper, we propose a fast face detection system to detect faces at any scale, angle and location in any images. To detect faces, under such mixed states, we introduce combination with a set of face re-normalize units and face or non-face classifiers. The face re normalize unit consists of three appearance-based (scale, rotation and translation) detectors, which have wider receptive field than that of the clarifiers. And the clarifiers classify faces after re-normalize procedure. Because of this wider receptive field, the re-normalize unit allows us sparse scanning steps on the image to reduce the number of candidates. However, they transform each candidate into canonical face precisely due to competitive dynamics of the detectors. We apply this system to face detection on CMU rotated image database. As a result, detection rate is 85.7% at 4 false positive per image. This detection rate is comparable to former work [Rowley,1998]. But the number of candidates after re-normalizing is 234,918 and it is less than 1/100 of the former work. Also, we apply this system to face detection on 150x220 pixel color image database. As a result, averaged detection time is 1.26 second per image. This is about 2 times as fast as former work [Hsu,2002].

Important Links:

Go Back