Learning and analysis of facial expression images using a five-layered hourglass-type neural network

In this study, a method to perform feature extraction and image creation support of five facial human expressions in gray scale images are presented in which the five-layered hourglass-type neural network is used. Input values to the neural network are five facial expression images composed of 100×100 pixels and are the same as the teacher-signals in the output layer. The teacher-signals are learned using the five-layered hourglass-type neural network to achieve a compression function and restoration of the images. The compressed information and causality of each facial expression are dealt with in the third layer (the emotion-layer). Furthermore, it can be shown that an image creation is performed by giving adequate values to the emotion-layer