A considerable number of studies have been made on the facial expression recognition techniques in pshycological field, engineering and so on. However, it remains an unsettled problem that trade-off between recognition accuracy and calculation cost. It disturbs realizing real-time processing. In the last few years, real-time recognition systems which use a high-speed graphics workstation or a transputer have been seen. They need such a high-end computer system, so it is difficult to use it as a simple interface between computer and human. In this paper, we propose a real-time facial expression recognition method on the assumption that it runs on generic (low-cost) work-station or PC with video capture function. A face extraction is based on simple temporal differential image. One-dimensional correlation matching method is used for a feature tracking. And performing discrete cosine transform (DCT) to image, calculated coefficients in term of festure vectors are given to the neural network. In the user depended expression classification experiments, we confirm that the accuracy of our method is above 90% for five expression categories.
抄録全体を表示