To provide a face recognition system capable of absorbing fluctuation of position/angle of a face or fluctuation of illumination and performing efficient learning and recognition even in case that the number of identification objects is increased.
A neural network NN is divided to a plurality of sub-nets, as shown in the drawing, to perform learning and recognition in parallel in sub-net units. In configuration of a parallel NN, it is permitted that the same facial pattern fuzzily attributes to a plurality of sub-nets, considering the ambiguity of patterns. According to this, relatively close patterns can be always efficiently recognized in the same sub-net. The facial pattern of an identification object is acquired as a there-dimensional image by a three-dimensional shape measuring device. Therefore, an input face pattern corrected while eliminating an influence such as a change in position or attribute of the person or a fluctuation of lighting as much as possible can be obtained, and a face recognition system robust to the rotation of a face is realized.
YAHAGI TAKASHI
SEKIYA DAIYU
CSNG200500057002, 呂建明、谷萩隆嗣, "ニューロファジィシステムによる顔画像認識の一方法", 電気学会論文誌C Vol.123 No.9 IEEJ, 20030909, 第123巻, p1555−1563, JP, (社)電気学会
JPN6010004323, 藤本稔、呂建明、関屋大雄、谷萩隆嗣, "3次元情報を利用した顔認識", FIT2002 情報科学技術フォーラム 一般講演論文集 第3分冊 画像認識・メディア理解 グラフィク, 20021005, p121−122, JP, 社団法人電子情報通信学会、社団法人情報処理学会
JPN6010004328, 呂建明、谷萩隆嗣, "ニューロファジィシステムによる顔画像認識の一方法", 電気学会論文誌C Vol.123 No.9 IEEJ, 20030909, 第123巻, p1555−1563, JP, (社)電気学会