Abstract
This paper proposes a new optimizing system for improved recognition of facial expressions based on a feature extraction filter which uses genetic programming (GP), Haar-like features and neural networks (NN). The purpose of the system is to gather data useful for the recognition of facial expressions using GP and Haar-like features and to evaluate this data's validity. The system employs a technique which converts images of facial expressions into feature images using Haar-like features, then extracts features from these converted images using GP. GP creates individuals using the feature-extracting filter and optimizes them. NN then classifies these facial expressions. The classification rate for the final generation of images was 0.88 for neutral faces, 0.6 for pleasure, 0.6 for sadness, 0.72 for surprise, 0.8 for anger, 0.6 for disgust and 0.44 for fear. The reason for the low classification rate for fear was that high fitness individuals could not be obtained. In addition, the system encountered difficulties in distinguishing images with very similar expressions, such as fear and disgust. On the other hand, compared to the previous system, the processing speed of the new system was markedly quicker. This new approach enables basic experiments to be repeated in a short-term cycle and makes it possible to search for important factors, thus realizing the full potential of the system. A number of practical applications for the system can be expected based on this new approach.
