Machine vision recognition of facial affect using backpropagation neural networks
R.R. Avent, Chong Teck Ng, J.A. Neal · 2002
A machine vision system has been developed to classify posed facial expressions indicative of eight discrete emotions: interest, happiness, sadness, surprise, anger, fear, contempt, and disgust. The system consists of an image capture subsystem to capture and store color images of facial affect; a face detection subsystem to distinguish image corresponding to a face from image corresponding to scene background; an edge detection subsystem to locate image regions corresponding to edges; a face feature detection subsystem to locate edge clusters corresponding to eyes, eye brows, and lips; and a face feature analysis subsystem consisting of eight backpropagation neural networks to classify edge dusters corresponding to facial features into the aforementioned emotion categories. An image database was created for system development and verification and consisted of facial affect captured from nine racially diverse males and females. System accuracy ranged from 68% to 89% across emotions and across the set of image resolutions utilized.