抄録
In order to develop "Active Human Interface, (AHI)" that realizes emotive communication between intelligent machine and human being, we have been investigating machine recognition of human emotions from facial expressions. Using a neural network (NN), we have already obtained a high correct recognition ratio of about 90[%] for 6 basic facial expressions, but we find that, depending on the NN learning information, there exist certain characteristics in the NN misrecognition and correct recognition. This paper analyzes these characteristics using action units (AUs) employed in the psychological field for describing the basic muscle movements of human face and analyzing the relationship between the NN input and output information. We clarify, in this paper, the reason why misrecognition occurs by using the Jacobian of the NN input-output function and how AUs influence the recognition results through sensitivity analysis done by assigning a zero value to each AU used in NN input information.
