05-05-2011, 04:32 PM
Abstract
This paper describes a fast face identification methodwith Gabor filters. Two efforts are made to achieve the acceptableprocessing speed: 1) we design the optimal Gaborfilters with the arrangement theory that uses a few directionsand layers. 2) The transformation with Gabor filters(as called Gabor transformation) is only done over the regionsaround the facial feature points, not the whole inputimage. The facial feature points extraction is performed bydetecting the facial organ regions with color informationand edge information, followed by the corner detection ineach detected facial organ region with the SUSAN operator
1 Introduction
Face identification is very important for security, surveillanceand telecommunication. Various approaches of faceidentification have been reported. Some of them are templatematching [1], eigen face method [2], etc. There arealso some methods using frequency information as featurevectors [3] [4] [5].Gabor transformation is one of the recognition methodsthat use frequency domain information. It can analyze informationabout both the spatial-domain and the frequencydomainsimultaneously with a signal. However, the highcomputation cost of Gabor transformation hinders its use inreal time image analysis, even with the help of FFT.In this paper, we propose a novel method to distinguishan individual in high speed by using optimal Gabor filters.In order to reducing the processing time, we design Gaborfilters with a filter arrangement theory [6]. The resultingGabor filters have only 4 directions and 3 layers, which aremuch simpler, compacter and more efficient. Another effortwe have done is that we only apply the Gabor transformationto the regions around the facial feature points, not thewhole input images. In the proposal method, we first detect face in an inputimage and estimate its pose by using color information.Then we rotate the input image according to the estimatedface pose to make the face upright. Next, we extract thefacial organs with both color and edge information. Afterthat, we merge the two eyes region into one and then applythe Gabor transformation to it. The two outer cornersof the two eyes are detected with an integral projection of aGabor filter output. The detected two outer corners of thetwo eyes are then used to normalize the size and the poseof the face. The facial features are detected by applying theSUSAN corner detector in operator in each extracted facialorgan region. Finally, we apply the Gabor transformation toeach small region around the detected facial feature pointsand perform the face identification by comparing the similarityof the Gabor filter outputs with ones of each registeredface.
Download full report
http://wakayama-u.ac.jp/~wuhy/ICPR02_Yoshida.pdf